Preprint
Review

Integrating Artificial Intelligence for Advancing Multiple Cancer Early Detection via Serum Biomarkers: A Narrative Review

Altmetrics

Downloads

129

Views

88

Comments

0

A peer-reviewed article of this preprint also exists.

Submitted:

31 December 2023

Posted:

03 January 2024

You are already at the latest version

Alerts
Abstract
The concept and policies of multiple early cancer detection (MCED) have gained significant attention from governments worldwide in recent years. In the era of burgeoning artificial intelligence (AI) technology, the integration of MCED with AI has become a prevailing trend, giving rise to a plethora of MCED AI products. However, due to the heterogeneity of both the detection targets and the AI technologies, the overall diversity of MCED AI products remains considerable. The types of detection targets encompass protein biomarkers, cell-free DNA, or combinations of these biomarkers. In the development of AI models, different model training approaches are employed, including datasets of case-control researches or real-world cancer screening datasets. Various validation techniques, such as cross-validation, location-wise validation, and time-wise validation are used. All of the factors show significant impacts on the predictive efficacy of MCED AIs. After the completion of AI model development, deploying the MCED AIs in clinical practice presents numerous challenges, including presenting the predictive reports, identifying the potential locations and types of tumor, and addressing cancer-related information such as clinical follow-up and treatment. This study reviews several mature MCED AI products currently available in the market, detecting their composing factors from serum biomarkers detection, MCED AIs training/validation, and the clinical application. The review illuminates the challenges encountered by existing MCED AI products across these stages, offering insights into the continued development and obstacles within the field of MCED AI.
Keywords: 
Subject: Medicine and Pharmacology  -   Oncology and Oncogenics

1. Introduction

1.1. Background and Motivation

Historically, cancer early detection has revolved around procedures like the Pap test introduced around 100 years ago (1). Early cancer detection is a pivotal factor in improving patient survival rates as it allows for earlier intervention, including the surgical removal of localized solid tumors, thus avoiding the critical metastatic stage where survival rates sharply decline to less than 50%, even with the most advanced systemic therapies (2). In many instances, cancer progresses over the course of years from its initial site to the point of metastasis, presenting a window of opportunity for early detection (3). The COVID-19 pandemic disrupted healthcare systems globally, leading to the postponement or cancellation of non-essential medical services, including cancer screenings. Recent increases in cancer-related mortality rates observed might partly be attributed to delays in cancer detection due to disrupted screening services (4).
Early cancer detection would improve prognosis, however, this conventional approach faces notable challenges. In fact, a substantial proportion (~57%) of cancer-related deaths in the United States are attributed to cancers that currently lack effective screening protocols (5). For decades, cancer early detection has focused on detection of individual cancer types . These methods include low-dose chest computed tomography (CT) for lung cancer, mammography for breast cancer, Pap smear for cervical cancer, and fecal occult blood tests for colorectal cancer (6). Although these screening methods have contributed significantly to the early detection of cancer, the cancer types that the methods screen represent only a fraction of the overall cancer landscape (7). Only 14% of cancer cases in the U.S. are diagnosed through recommended screening tests, highlighting the limitations of traditional screening methods. The majority of cancer diagnoses occur after symptoms have manifested or during unrelated medical visits (8). Moreover, cancer screening by the individual cancer type-manner is inconvenient and individuals may find themselves required to visit multiple healthcare providers to undergo various screening tests. These drawbacks contribute to a reduced adherence to cancer screening through these methods (9).
The inadequacy of current cancer screening strategies can be attributed to several factors. These include suboptimal adherence to screening guidelines, disparities in access to screening, and, to a lesser extent, limitations in existing screening technologies and the occurrence of cancers between recommended testing windows (10). Consequently, a significant portion of cancer cases in the United States are detected at advanced stages, where treatment becomes substantially more challenging. For instance, among cancers with established screening protocols, the proportion of late-stage diagnoses ranges from 20.9% for prostate cancer to a staggering 64.7% for lung cancer.
This introduction sets the foundation for exploring the promise of Multi-Cancer Early Detection (MCED) strategies, which aim to revolutionize cancer screening by addressing these challenges. Subsequent sections will delve into advances in MCED approaches and the pivotal role played by Artificial Intelligence (AI) in enhancing the early detection of a broad spectrum of cancers, ultimately improving survival rates and treatment outcomes.

1.2. Role of Artificial Intelligence in MCED

Artificial intelligence (AI) is good at detecting hidden patterns within complicated datasets. Typically, supervised machine learning algorithms are the workable AI approach in the biomedical field. Because of its outstanding performance in classification or prediction, AI has been widely applied across multiple biomedical fields in recent years. The applications include correlating genetic data to obesity (11), liquid biopsy for predicting cancer metastasis (12), clinicopathological data for risk-stratification of cancer (13), genetic data for MCED (14), and protein biomarkers for MCED (15).
In the field of MCED, applying AI has also proven to be efficient in improving diagnostic performance (16). Harnessing AI has become a must-use technology in analyzing MCED data because MCED tools typically target tens of analytical targets. Interpretation of the complicated patterns that are composed of tens of components would be difficult without AI approaches. Even when some biomarker experts are good at interpreting complicated data patterns, the interpretation job is still time-consuming and labor-intensive. However, the diagnostic performance of using AI in MCED is still discrepant between different studies. There are many possible reasons for inconsistent results. One significant factor is the variation in study designs. Most related studies use data from case-control studies for training and validation of AI models (15,17). Only a few research labs use data from cancer screening for the development and validation of related AI models (6,18). Using real-world cancer screening data can be quite important because it is the only way to align with the real-world application of these models. AI models are data-driven and would be heavily influenced by the composition of training data. If the data used for training the model differs from the data of intended-use scenarios, then the use of AI models in real-world healthcare would be significantly limited. Analytical variation between different ethnic groups would also be another factor that must be considered. Only when factors that affect AI, such as input analytical data, model training/validation, are well-optimized and standardized, can the clinical field fully leverage the capabilities and advantages of AI in cancer screening.

2. Trajectory of Early Cancer Detection Methods

2.1. Evolutionary Overview of Multiple Cancer Early Detection

In recent years, significant progress has been made in the realm of cancer screening, especially in the domain of cancer early detection (19). Several single cancer screening tools, such as low-dose chest computed tomography (CT), mammography, pap smear, and colonoscopy, are currently applied for individual cancers. The evolution of MCED has been driven by the recognition of the limitations inherent in traditional single-screening approaches (19). In contrast to multimodal single cancer screening methods, MCED tests aggregate the prevalence of various cancer types within a given population (20). This approach provides a single, all-encompassing evaluation while maintaining a relatively low false-positive rate (20). Moreover, the challenge of varying levels of adherence to current screening protocols has complicated cancer detection (21), necessitating the development of non-invasive, multi-cancer screening methods to reduce morbidity and mortality from cancers.
Historically, the exploration of whole-body imaging and endoscopic techniques has been considered a route to achieving universal cancer screening (19). Nevertheless, persistent issues, such as high false positive rates (22) and potential complications arising from radiation exposure (22,23) or invasive procedures (23), still require resolution (24). In recent times, a revolutionary breakthrough has emerged in the form of liquid biopsies, which analyze cancer-related biomarkers present in body fluids (24). This development has introduced a transformative dimension to the field, offering a less invasive and more accessible means of early cancer detection. Additionally, AI has been recruited to analyze large amounts of data, including medical images and genetic data, to identify patterns and anomalies indicative of cancer (24). The incorporation of AI into the screening process increases the accuracy and efficiency of cancer detection. Of note in the current clinical workflow, MCEDs can serve as a precursor to more specific cancer diagnoses. Specifically, MCEDs should not be viewed as diagnostic but predictive for risk and these approaches inherently alter the acceptable level of specificity.

2.2. Advancements in Imaging and Endoscopic Tools

Imaging techniques, such as CT, magnetic resonance imaging (MRI), and positron emission tomography (PET) have potential in MCED by providing non-invasive methods to identify tumors in asymptomatic patients. However, these techniques have limitations and challenges. Imaging will always be limited to a minimal size of a lesion that is noticeable (even by AI) on a scan. False positives occur when imaging identifies benign abnormalities as suspicious, leading to unnecessary tests, interventions, and mental stress for the patient (22,25). For example, previous studies about whole body-MRI showed that abnormal findings are expected in about 95% of screened subjects, about 30% of subjects would require further investigations but less than 2% would be reported as suspicious for malignant cancers (22). Moreover, there is a carcinogenic risk associated with radiation exposure from these examinations (26). Annual CT scans from ages 45 to 75 years could result in an increased risk of cancer mortality of 1.9%, or approximately 1 in 50 people (23).
Endoscopic techniques, on the other hand, allow direct visualization and biopsy of suspicious lesions. The development of novel endoscopic technologies such as narrow-band imaging (NBI) and confocal laser endomicroscopy (CLE) has enhanced the ability to detect lesions in the gastrointestinal tract and other organs (27). However, these invasive procedures lack cost-effectiveness and may carry risks such as bleeding and bowel perforation. While the reported post-colonoscopy perforation rate is less than 0.1%, it remains a significant concern due to its status as a severe complication associated with high mortality rates (28).

2.3. Emergence of Liquid Biopsy-Based Approaches

Liquid biopsy has emerged as a revolutionary technique for MCED. This innovative method through phlebotomy greatly reduces the possible harm associated with more invasive screening methods. Liquid biopsy involves the analysis of disease-related markers found in bodily fluids, encompassing a diverse range of analytes such as circulating tumor DNA (ctDNA), circulating tumor RNA (ctRNA), circulating tumor cells (CTCs), proteins, and metabolites (24).
Over the past decade, there has been a rapid development and adoption of Next-Generation Sequencing (NGS)-based methods in cancer research (29). These methods allowed us to capture tumor-specific genomic aberrations in circulation (29). There are two primary sources of tumor DNA that can be noninvasively assessed within the circulatory system: ctDNA and CTCs (30). CtDNA consists of small nucleic acid fragments shed from necrotic or apoptotic tumor cells (31). In contrast, CTCs represent intact and often viable cells, which may originate from active cell invasion or the passive shedding of tumor cell clusters (30). Genomic biomarkers hold the potential to provide a more representative ‘summary’ of tumor heterogeneity within a patient and also open up the possibility of detecting cancer at an early stage (29). Several commercial products, such as GRAIL, have demonstrated impressive performance in detecting multiple cancer types and identifying their origin within asymptomatic patients (32).
On the other hand, serum protein tumor markers like CEA, AFP, CA-125, CA-19.9, PSA, and others have been used for decades to aid in diagnosing and managing various cancers (33). However, due to their relatively low sensitivity and specificity for early cancer detection, most international guidelines recommend their use primarily for monitoring cancer recurrence or assessing therapy response, rather than as screening tools for early detection (6). One potential strategy to address this limitation involves combining multiple serum markers into diagnostic biomarker panels (15,33,34). Previous research has shown that when AI algorithms are employed to train these serum marker panels, the resulting algorithms become effective tools for cancer screening (16). These AI algorithms consistently exhibit high levels of accuracy, generalizability, and cost-effectiveness, making them promising candidates for improving early cancer detection (6,35).

3. Serum Biomarkers as Critical Indicators

3.1. Protein Biomarkers: Unveiling Diagnostic Potential

Cancer cells or other cell types in tumor microenvironment release soluble molecules that are identified as serum tumor markers by non-invasive diagnostic assay. These molecules ideally detect disease early, predict response and aid in monitoring therapies. For example in breast cancer, different serum markers are carcinoembryonic antigen (CEA), the soluble form of MUC-1 protein (CA15-3), circulating cytokeratins such as tissue polypeptide antigen (TPA), tissue polypeptide specific antigen (TPS) and cytokeratin 19 fragment (CYFRA 21-1), and the proteolytically cleaved ectodomain of the human epidermal growth factor receptor 2 (s-HER2). These markers are used majorly in following-up (12), but are not used in screening breast cancer (36).
Protein tumor markers have not been fully exploited clinically both diagnostically and prognostically. Therefore, expansion from individual protein biomarker analysis to protein panels or proteomes, develops a comprehensive prognostic analysis to predict disease onset and progression (37,38). The protein panel analysis far exceeds the single biomarker analysis in facilitating specific intervention or guiding treatment especially in drug resistance. Challenges prevail in the transition from single biomarkers to proteomic panels both on the basis of process development and technicality. However, recent advancements in the proteomic techniques have fortified that analysis of multiple proteins simultaneously in the blood, urine, cerebrospinal fluid or any other biological sample (38).
The technical difficulties in tumor marker measurement include errors due to difference between labs and also within batches. These variation combinations to form a panel result in low robustness and reproducibility. Hence in the development of a robust panel assay over time and across laboratories, a single analytical parameter determined by a single method permits quantification of errors and batch variability. Further, results are compared by absolute quantitative technologies rather than relative quantitative techniques. Absolute quantification requires lack of dependency on affinity reagents instead are directed by mass spectrometry-based proteomics (39). The US FDA has approved 15 protein biomarker assays in serum and/or plasma . Of the 15 FDA approved protein biomarkers for cancer proteins, 9 are applicable for serum and 6 for serum/plasma. Although both plasma and serum are identical in protein composition, expression or recovery of individual proteins vary greatly. For instance, the free PSA concentration differs in serum and plasma (40). The HUman Proteome Organization recommends plasma for proteomics studies (41).
The idea of panel testing for the proteomic profiling has emerged as an effective method in diagnostics of cancer, particularly cancer proteomics is clinically feasible. The enzyme-linked immunosorbent assay, immunohistochemistry and flow cytometry system are reliable, sensitive and widely used in the clinical diagnosis, prognosis and treatment monitoring of cancer (42). Alternative techniques like mass spectrometry, protein array and microfluidics are extensively used and are being developed for clinical application (43). On top of the massive data created by panel testing, proteomic workflows for targeted analysis of protein panels have improved with highly standardized sample-preparation protocols (44), data-independent acquisition techniques (38), sensitivity and faster mass spectrometers conjoined with micro-, and analytical flow rate chromatography (45). The absolute quantification has improved the statistical analysis, cross-study and cross-laboratory comparability simplifying accreditation of analytical tests (46).
In 2009, OVA1 was approved for the evaluation of ovarian tumors in combination with the measurement of five serum proteins: apolipoprotein A1, β-2 microglobulin, CA -125, transferrin and transthyretin (47). In 2011, ROMA was approved for the prediction of ovarian malignancy along with two proteins - human epididymis protein 4 and CA-125 (48). For early detection of cancer, a total of 1261 proteins were identified that were involved in oncogenesis; tumor - angiogenesis, differentiation, proliferation, apoptosis; in cell cycle and signaling. In as many as 1261 proteins, 9 protein biomarkers have been approved as "tumor associated antigens" by the USFDA. Although these protein biomarkers have not yet been approved for MCED, in many Asian regions such as China (35), Taiwan (33), Korea (49), etc., the use of protein biomarkers for MCED has been put in practice for more than 10 years. The popularity of this approach lies in its convenience, as cancer screening for many different cancer types can be conducted with a simple blood test. This includes many cancer types for which there is no preferred screening method (18). Additionally, the cost of protein tumor marker tests is relatively low; the cost of one marker test may be around $10 USD or even lower, making it financially feasible for widespread use. In terms of the diagnostic performance, using protein biomarker panels can achieve approximately 40% sensitivity and 90% specificity (33). In regions with high accessibility of follow-up diagnostic approaches (e.g. endoscope, CT, MRI), this is a convenient and competitive approach.

3.2. cfDNA Biomarkers: Unleashing Genomic Clues

Cell-free DNA (cfDNA) are non-invasive markers detected in serum, plasma, urine and CSF (50) and a more favored biomarker for cancer surpassing the gold standard approach of biopsy sampling which is invasive with restricted frequency of usage and site. It depicts tumor heterogeneity with a comprehensive representation allowing multiple sampling from a single blood draw to and represents various tumor clones and sites providing a comprehensive representation (51). All cells release cfDNA that may be necrotic or apoptotic. The cfDNA reveals mutations, methylation, and copy number variations that may be related to cancer (52). Hence, its molecular profiling has a potential role in non-invasive cancer management with the advent of ultrasensitive technologies (e.g., NGS, BEAMing (beads, emulsions, amplification and magnetics), and droplet digital PCR (ddPCR)). It has evolved as a considerable surrogate marker in tumor detection, staging, prognosis, localization and in identification of acquired drug resistance mechanisms (53).
The sensitivity to detect tumor derived cfDNA is expressed in terms of mutant allele fraction (MAF) which is the ratio between the amounts of mutant alleles versus wild-type alleles in a sample. The MAF detection limits of quantitative PCR ranges between 10%–20%. However, variations of PCR techniques like allele-specific amplification (54), allele-specific non-extendable primer blocker PCR (55), and peptide nucleic acid-locked nucleic acid PCR clamp (56) increase the sensitivity. Several genome-wide sequencing methods have been developed in the last decade. The methods include Plasma-Seq (57), Parallel Analysis of RNA ends sequencing (58), and modified fast aneuploidy screening test-sequencing (59) for cfDNA detection at 5–10% MAF. Targeted sequencing approaches include the exome sequencing (60), CAncer Personalized Profiling by deep Sequencing (CAPP-Seq) (61), and digital sequencing (62). Targeted sequencing approaches are high coverage whereas whole genome sequencing (WGS) approaches are of low coverage. Targeted approaches detect mutations even at low ctDNA, whereas WGS assess copy number alteration in ctDNA. A lower MAF is obtained with digital PCR (dPCR) method, including microfluidic-based droplet digital PCR (ddPCR) and BEAMing (63) quantified with extreme sensitivity (0.001%-0.05% MAF). The multiplexing capabilities are limited as the primers or probes target specific mutations or loci.
For the purpose of MCED, cfDNA detects a tumor at an asymptomatic stage with a diameter of 5mm. The ratio of tumor-derived cfDNA to normal cfDNA < 1–100,000 copies (MAF of 0.001%) corresponds to a tumor of 5mm diameter (64). In blood, 1 mL of plasma contains approximately 3000 whole-genome equivalents (65) and in the total 3L plasma represents 9,000,000 copies. In the entire cfDNA population, only one cancer genome originates from a 1 mm diameter tumor, increasing the probability of extracting one tumor-derived cfDNA fragment from a 10 mL blood sample which is very low. Hence, these available methods detect tumors with a diameter greater than 1 cm (0.5 cm3) (64). Different from protein-based methods, tumor-derived cfDNA are DNA fragments released from dying cancer cells and DNA copy numbers are limited in a cell. Thus there is a limit of detection and a potential limit to how early detection can occur. Thus, if a cancer-associated MAF is detected, it is likely cancer. Protein biomarkers are released by cancer cells at a relatively high amount so easily detectable early (39,66,67), but lack specificity because protein biomarkers can be released by both cancer cells as well as normal cells.
The cost of cfDNA testing has significantly decreased in recent years, although it is still over five times the price of protein biomarker panels (5). However, it can generally be achieved at a cost below $1,000 USD. The price reduction may lead to increased accessibility, however, there are still some inherent issues with cfDNA testing that remain unresolved. One critical concern is its short half-life, potentially as brief as a few minutes to hours (68). Such a short half-life would result in an unstable cfDNA quantity in the specimen. Additionally, specimen preservation would pose a challenge, as the cfDNA could degrade within a few hours of in vitro storage. In contrast, protein biomarkers have a half-life lasting several days or even weeks (69,70). These inherent issues may be reasons why the effectiveness of cfDNA testing in MCED is not as promising as initially anticipated. In fact, a study suggests that combining cfDNA with protein biomarker testing does not yield better cancer efficacy than using protein biomarkers alone (71). Further optimization is required for the use of cfDNA testing in MCED.

4. Synergizing AI Algorithms for Biomarker Analysis

4.1. Classical Machine Learning Techniques in Biomarker Interpretation

Harnessing ML in interpreting clinical inputs for classification or prediction is becoming a mainstream application nowadays in the medical field. Several studies have indicated that ML algorithms analyzing clinical (72), genetic (13), or protein biomarker (16) results can provide diagnoses similar or even better than those made by physicians . What is noteworthy is that ML algorithms demonstrate greater consistency in pattern recognition, reducing inter-individual differences. In the medical domain, there exists a wide variety of ML algorithms, including logistic regression, decision trees, random forests, support vector machines, and more (73,74). Despite differences in the underlying logic of these algorithms, their design aims to identify specific patterns and relationships between the data and the predicted targets.
In cancer screening or diagnosis studies, the effectiveness of ML algorithms was compared with physician interpretation of tests (16,75). In these studies, human physicians used the reference range-based single threshold method: predicting the probability of cancer occurrence within the next year if any test item exceeded the reference range. Conversely, if all test items fall within the reference ranges, the individuals are predicted not to be at risk of cancer. While this interpretation method is straightforward, the effectiveness of cancer screening is not as high as that achieved by machine learning algorithms. This suggests that physicians may not be as sensitive to specific data patterns in laboratory test results as ML algorithms. The possible explanation is that ML algorithms detect the “face/pattern of a disease” rather than only a few test items.
While ML algorithms appear to generally outperform physicians in interpreting multiple test items, there doesn't seem to be a particular advantage among different ML algorithms for lab data-based classification problems in the medical field. Although in individual reports, various algorithms like support vector machine (75), random forest (76), logistic regression (16) have been reported to outperform others. In a review study, it was also noted that other ML algorithms do not show a clear superiority over traditional logistic regression (also categorized as a ML algorithm) (77). In fact, most MCED products still adopt logistic regression as the ML algorithm. Galleri, a cfDNA based MCED AI is composed of two logistic regression models, one for cancer detection and the other is for predicting tissue of origin (78). Protein biomarkers based MCED products such as OneTest (6) (20/20 GeneSystems) and CancerSEEK (Exact Sciences) (15) also revealed the utility of the classical ML algorithms. Overall, despite some ML algorithms seeming more prominent in these studies, their advantages are very limited. In fact, the nature of the laboratory data itself determines whether such classification problems have good predictive performance. The data has already predetermined the predictive performance, and the choice of which ML algorithm to use doesn't play a significant role (79).
The reason why data predetermines outcomes can be explained that the lab data-based AI models are based on lab data and the lab tests typically have a good signal-to-noise ratio (77). Moreover, these test items have undergone a series of rigorous validations from the development stage and implemented in clinical settings for years (80). Thus, the lab test items fundamentally have a certain correlation with the predictive phenotypes or diseases. On top of that, tests like proteomic panels consisting of peptides and proteins would not suffice as biomarkers on their own but instead acquiring a ML strategy for their interpretation renders good predictive performance (81). On the basis that the data itself is composed of such strong predictors, ML models built on either theoretical foundations can easily identify hidden patterns in the data. In summary, for medical AI models with lab data as input, the importance of good data far outweighs the significance of the ML algorithm used.

4.2. Unveiling Deep Learning's Potential in Biomarker Analysis

In recent years, deep learning (DL) algorithms have achieved significant success in the field of computer vision. In the domain of medical imaging, DL algorithms are widely employed for the development of image recognition models. Medical images such as electrocardiograms, chest X-rays, and computed tomography scans are particularly well-suited for the application of DL algorithms. In these areas, DL algorithms demonstrate excellent performance, often approaching the level of human experts (82). One key distinction between DL algorithms and traditional ML algorithms lies in feature engineering. Typically, when dealing with high-dimensional data, traditional ML algorithms require the use of feature extraction or feature selection methods to reduce the data dimensions in order to improve prediction accuracy. In contrast to traditional ML algorithms, DL does not necessitate upfront feature engineering (83). Therefore, DL offers the convenience of not requiring these preprocessing steps over traditional ML and provides a distinct advantage in practice.
While DL algorithms have achieved significant success, it appears that they do not necessarily outperform conventional ML algorithms in the medical domain. For instance, traditional ML- random forest method attained higher diagnostic performance than DL in ultrasound breast lesions classification (84). In a study predicting postoperative patient conditions, DL algorithms did not demonstrate higher predictive capabilities compared to traditional logistic regression (85). In another study predicting drug resistance based on mass spectrometry data, random forest or XGBoost algorithms exhibited higher predictive abilities than DL (80). In the field of MCED, models using DL algorithms to analyze protein biomarker results did not show higher cancer prediction capabilities than traditional ML algorithms like logistic regression (35). In certain data structures where the data itself contains strong predictors, the need for feature engineering in DL algorithms is not as apparent as in traditional ML algorithms (80). Thus, The performance comparison between DL and ML depends significantly on the data structure inherent to the specific application (84).
In situations where there is no advantage in predictive performance, the use of DL algorithms to analyze lab test results becomes debatable. Due to the complex computations within the model, DL algorithms require more processing time to generate classification or prediction results (80). Beyond the longer processing time, DL algorithms also consume more energy compared to traditional ML algorithms to produce predictive outcomes (80). In an era where AI algorithms are gradually becoming a part of daily life, energy-intensive methods pose a higher carbon footprint, eventually facing serious challenges. While there may not be a significant advantage in predictive performance, certain DL algorithms can assist in addressing clinical challenges encountered in the real world. Taking the field of predicting cancer risk using protein biomarkers as an example, the test panel provided by each diagnostic institution may vary, with only partial overlap in the panels tested. Additionally, if the items tested for each case only partially overlap at different time points, comparing risk predictions becomes challenging. In this regard, long short-term memory networks, with their flexibility and tolerance for missing values, prove to be a suitable solution for addressing such clinical issues (35).

5. Training and Validation of AI Models for MCED

5.1. Impacts of training dataset: case-control, retrospective cohort, or prospective cohort?

AI technology is very promising for many applications in medical fields (11,72,86). However, the robustness of the medical AI models is suboptimal when deployed into real world settings (87). The suboptimal robustness indicates that the medical AI models perform well in training and validation, but such models fail to perform in a real world deployment. While the underlying mechanisms are many, for MCED AI model, the most crucial factor for a suboptimal performance would be resulted from inadequate selection of training datasets. The training dataset types can be categorized into case-control cohort, retrospective cohort, and prospective cohort (Table 1). The difference on the training dataset determines what the MCED AI models learn. Basically, MCED AI would perform well when the training datasets mimic the real world settings. By contrast, diagnostic performance of MCED AI models would drop if the training datasets are considerably different from the real world settings. Typically, training by dataset with case-control design is the most susceptible to failure when deployed in the real world, even the training datasets are reasonably designed according to classical principles of ML models training. In training a ML model of binary classification, the training datasets include cases with positive label and cases with negative label. For MCED, the positive label indicates positive for cancer diagnosis while the negative label indicates cancer negative (ie., healthy cases). Cancer cases and healthy cases are the only learning materials for ML algorithms. Appropriate datasets are the key to successful and useful MCED models. By contrast, inappropriate datasets would lead to disastrous deployment even though the models perform well in the training processes.
The differences between a case-control study and a real-world cohort study for cancer screening would be as follows:
  • On cancer cases: Specimens of the cancer cases in a case-control study are typically collected in more advanced stages than the specimens of a real-world cohort study. The reason for that is that the specimens of the cancer cases in a case-control study are collected when the diagnosis of cancer has been made, which is often associated with symptoms/signs that are caused by cancers. In this case, the cancers show their malignant behaviors like space occupying or mass effect. By contrast, specimens of the cancer cases in a real-world cohort are collected long before cancer diagnosis or any symptom/sign. Such conditions are usually closer to the health check-ups population in the real-world. Theoretically, biomarkers in presymptomatic or asymptomatic cancer cases would be closer to those of healthy controls than in the symptomatic cancer cases.
  • On healthy cases: The number of healthy control cases in a case-control study is usually up to several hundred given the fact that the ratio of cancer versus control ratio is set around 1:1-1:4 (15,17,34,49). The relatively small number cannot represent the large diversity in the healthy control cases. As a result, there are fewer outlier cases. Fewer healthy outliers would simplify the classification problem (i.e. classify cancers versus healthy). AI models trained with fewer healthy outliers may therefore not have a classification threshold that can be used in the real world.
Figure 1 illustrates the probability distributions of the AI models trained by case-control dataset or real-world dataset. Consider a case-control study with ratio of cancer to health cases = 100:100 and a real-world cohort study with ratio of cancer to healthy cases = 100:10,000. In the training and validation steps, AI models in the case-control study can easily reach a good performance when the cutoff of risk score is in the range 10.49-15.49 (Figure 1A). By contrast, optimal cutoff would be largely converged to a smaller range or a number (e.g. 15 in the example, Figure 1B). As illustrated in the plots, the cutoff of AI risk scores in the case-control study are not optimized for use in the real world. Many models (or many cutoffs) can have good classification performance. However, once in the real world, the performance of these AI models will be significantly weakened or changed. For example, when an AI model with Cutoff.1 (i.e. 10.49) is used in a real-world setting, the specificity would largely decrease; when the Cutoff.2 (i.e. 15.49) is used, the sensitivity would largely decrease. Of note, AI models with either Cutoff.1 or Cutoff.2 attain nearly 100% for sensitivity or specificity, but fail to deliver similar performance.
Failure of AI models that were trained from case-control studies in a real-world setting can be explained by t-value theory. According to the t-value formula, the disparity of means relative to variances between the two populations (cancer group vs. healthy group) determines the significance of the difference between populations (88). For a case-control study, a larger difference between the averages of the two populations indicates greater statistical significance between the two populations and renders it easier to be classified. Moreover, MCED AIs that are trained on case-control will not recognize the smaller differences in marker values that will be seen in early diagnosis because such asymptomatic/presymptomatic cases are missing in such datasets. Thus, the sensitivity of a case-control trained AI model will be greatly lost in an asymptomatic early population. Furthermore, it may rely more on markers that dominate later in the cancer development process.

5.2. Cross-Validation vs. Independent Testing: generalizability or continual monitoring matters?

In the general development of AI models, validating the performance of the model is a crucial step. There are various methodologies for model validation, including k-fold cross-validation (KFCV), nested k-fold cross-validation (NKFCV) , and independent testing. KFCV is typically used for initial internal validation, meaning it uses a single-source dataset while developing and validating the model. NKFCV is employed for smaller datasets, and it involves placing the model-tuning steps in a separate inner layer to avoid overfitting. Independent testing is commonly considered the fairest method in the machine learning training and validation process.
For MCED AIs trained by real-world datasets, special data processing is necessary to cope with the extremely balanced data structure, in which cancer versus non-cancer ratio is around 1:100 (16). To better train AI models, typically the number of cases with positive label (i.e. cancer) and the number of cases with negative label (i.e. non-cancer) should be comparable. Over-sampling of cancer cases or under-sampling of non-cancer cases can be adopted to create a balanced dataset for model training (Figure 2A). Under-sampling of non-cancer cases would be a more acceptable method in the medical community because no artificial data are created. In the model training step, stratified sampling rather than simple random sampling is adopted to divide the data into training dataset and validation dataset (Figure 2B). With stratified sampling, case number of the minor subgroup (i.e. cancer cases) is elevated so there will be sufficient cancer cases in the training dataset (89). Of note, the cancer versus non-cancer ratio should be kept the same as that of the original real-world cancer screening dataset in the validation dataset as well as in the independent dataset. Only by keeping the original cancer versus non-cancer ratio so that the static metrics of AI models can be accurately estimated.
Independent testing can further be divided into using data from a second or even a third healthcare institution to validate the AI model trained on data from the first institution (Figure 2C). This approach ensures that the AI model trained by the first institution is not limited to its specific context but can be applied more broadly. This characteristic is referred to as generalizability and typically regarded as one of key characteristics for AI models (90). However, should the use of medical AI be predicated on its generality? This issue has raised considerable interests in recent years. Medical behaviors are significantly influenced by the local socioeconomic status and healthcare insurance, coverage, as well as reimbursement. In places where healthcare costs are low, and accessibility is high (such as in East Asia), people are accustomed to undergoing annual cancer screenings, even without any symptoms (16). Conversely, in places where cancer screening is expensive and highly inconvenient, individuals might only opt for screening when experiencing significant symptoms or discomfort. Therefore, despite both being considered cancer screening databases, there is a substantial difference between them: categorized into presymptomatic/asymptomatic or symptomatic cases. These two groups exhibit significant differences in the progression of the disease, but current research has paid less attention to this aspect. This phenomenon creates a challenge for AI models for cancer screening, as obtaining similar results across datasets from different locations is difficult due to the inherent differences in these datasets.
Because data collected from different places exhibit considerable heterogeneity, and there are significant variations in healthcare systems and reimbursement structures. Therefore, pursuing the generality of AI models for cancer screening would contribute minimally to increase the robustness of local medical service. In contrast, time-wise management appears to be a more locally-relevant strategy for AI-driven MCED (Figure 2C). Before the AI models provide services, it's essential to verify whether the predictive accuracy remains stable across different years. Additionally, offering services locally and continuously, recurrently monitoring the model's performance over time helps ensure its stability. This involves training the model with local data and testing it with local data, aiming to provide better local healthcare services. The primary focus of local healthcare should be to serve the local community effectively, and there seems to be little benefit in overly pursuing generalization. Although this approach deviates from the conventional emphasis on the generalization of AI models, it maximizes healthcare benefits.

6. Challenges and Opportunities

6.1. Data Quality and Quantity: Navigating the Complex Landscape

While the performance of medical AIs depends largely on data, the characteristics of data would serve as the cornerstone for the performance and applications of MCED AI models. There are still several key challenges ahead and careful consideration needed. One of the foremost challenges revolves around the inherent heterogeneity in datasets sourced from various medical institutions and laboratories. The reasons for such variations are manifold. Firstly, differences arise at the level of analytical measurements, where even for the same biomarker, various laboratories and medical facilities may employ reagents or platforms from different suppliers. Studies indicate that variations would exist between assays for the same biomarker on different analytical kits, emphasizing the need for extensive and rigorous validation before clinical implementation (91). Additionally, discrepancies in the biomarkers that are tested in the panel also pose computational challenges and interpretation issues for AI models. Taking the example of protein biomarkers, one MCED screening kit may assess seven protein targets (16), while another MCED AI examines 12 protein targets (49). Yet another screening kit may include both protein and gene targets (15). While some items overlap among these kits, the differences would complicate interpretation and comparison. In this case, there is often an absence of comprehensive input data during the computation of AI models . To address the missing values, reliance on various data imputation methods becomes necessary. The methods for data imputation are diverse (92), but the imputed values generated through these methods may not represent real-world values. This imputation stage introduces some biases into the calculations. Moreover, regional or national heterogeneity also plays a significant role. Beyond factors related to diverse ethnicities, variations in health insurance reimbursement systems and healthcare cultures across different regions or countries can profoundly impact data collection (93,94).
Acquiring a training dataset that can represent the real-world is crucial for a MCED AI model to be effectively used in cancer screening. Ideally, such locally-relevant data should be continuously collected without selection from local routine medical practices. However, collecting such datasets poses significant challenges due to the scarcity of cancer cases compared to healthy cases (95). An extremely imbalanced dataset is inevitable in the real world setting (Figure 2A). The primary difficulty is the inadequate collection of a sufficient number of cancer cases, especially for rarer cancer types. Under the supervised learning training framework, insufficient positive cases will prevent the training of a trustworthy AI model. Therefore, collecting a substantial number of cases is absolutely critical. Unfortunately, in the current framework, patients or healthcare institutes lack sufficient incentives to contribute data because there are no additional rewards, and there are concerns about privacy and security of personal information (96). To address this dilemma, apart from some open-source medical datasets (15,97), there are emerging blockchain-based technologies that incentivize patients to upload medical data (98). These technologies not only reward patients for contributing their data but also create a cyclical reward system for the AI model's profits, all while preserving individual privacy through blockchain technology.

6.2. Interpretability, Explainability, and integration: Bridging the Gap in AI-Driven Insights

In addition to the issues related to data and AI algorithms, the most significant challenge in implementing a MCED AI model in clinical applications is likely how to interpret and communicate the results generated by the AI models (Figure 3). With extensive research in recent years, AI models in healthcare have made significant progress. However, beyond predictive performance, correctly explaining the reasons for such predictions is crucial to gaining the trust of clinical physicians and patients. Information such as the coefficient in logistic regression or RF Importance in random forests can provide rationale for predictions. Only when clinical physicians can clearly explain the reasons for the model's interpretation can they trust that the AI model's interpretation is reasonable and not just a "Black box." Establishing such a trust relationship between clinical physicians and AI models is a crucial first step for the success of MCED AIs. Moreover, clinical factors like age or sex are informative and crucial inputs for clinical diagnosis. The importance of such clinical factors would be as important as the serum biomarkers (16). While these clinical factors are the typical factors to be considered in the clinical diagnosis, including more clinical factors together with the serum biomarkers in MCED AI models would provide a more comprehensive solution to healthcare professionals.
On top of a reasonable and explainable MCDE AI model, the next crucial question that physicians would be interested in is the clinical relevance and actionability (99). To communicate with patients, the MCED AIs should provide more communicable terms for the predictive results. The predictive probability generated by AI models may not be an adequate metric for communication with patients. Instead, the incidence-based risk score (or positive predictive score, PPV score) that has been widely used in prenatal checkups is a more appropriate term (100). The PPV score is based on comparing the patient’s predictive probability to the cases with similar risk levels. For example, risk of 1 in 10 is interpreted higher than background risk whose risk level is 1 in 1,000. The population-derived PPV score would be more intuitive for non-medical professionals. Besides, PPV score is also an explicit metric rather than simply predictive probability to have comparison to background risk.
Unfortunately, only a few MCED AI models provide actionability to clinical physicians and patients, leaving a considerable gap between a report of elevated risk and the following cancer diagnosis. For cases with elevated risk for cancers, the next diagnosis of interest is the staging and localization. Staging relevant to the risk score would provide earlier information for prompt action (6). Localization is also a key information for MCED AI models. When tissue of origin is provided together with the risk score (6,15,18), physicians would know which medical specialty to suggest to the patients with elevated risk score. Actionable suggestions following a MCED test should be also provided to complete the whole test-and-action cycle. The call-to-actions that are built based on evidence can include the time interval for following-up, re-testing, and visiting to medical subspecialties (6,18). All of the actions can be depicted in a flowchart for easy guidance for MCED-using physicians so that a MCED AI can be seamlessly integrated into clinical workflows, facilitating informed decision-making by healthcare professionals.

7. Conclusion

Serum biomarkers-based AIs hold promise in MCED and are undergoing rapid development. The analytical targets include cfDNA, protein biomarkers, or their combination. When the serum biomarkers typically have the characteristics of strong predictors, various ML algorithms can have good diagnostic performance. The technical key of building a trustworthy MCED AI resides in using real world-mimatic data rather than a case-control design for training and validation so as to have a robust implementation back in a real world setting. Like other medical products, MCED AIs with high interpretability, explainability, and actionability would integrate better into medical workflow and benefit more patients in early cancer diagnosis.

References

  1. Zutshi V, Kaur G. Remembering George Papanicolaou: A Revolutionary Who Invented the Pap Smear Test. J Colposc Low Genit Tract Pathol. 2023 Aug;1(2):47. [CrossRef]
  2. Siegel RL, Miller KD, Fuchs HE, Jemal A. Cancer statistics, 2022. CA Cancer J Clin. 2022;72(1):7–33.
  3. Vogelstein B, Kinzler KW. The Path to Cancer — Three Strikes and You’re Out. N Engl J Med. 2015 Nov 12;373(20):1895–8.
  4. Fedeli U, Barbiellini Amidei C, Han X, Jemal A. Changes in cancer-related mortality during the COVID-19 pandemic in the United States. JNCI J Natl Cancer Inst. 2023 Sep 9;djad191. [CrossRef]
  5. Guerra CE, Sharma PV, Castillo BS. Multi-Cancer Early Detection: The New Frontier in Cancer Early Detection. Annu Rev Med. 2024;75(1):null. [CrossRef]
  6. Wang HY, Chen CH, Shi S, Chung CR, Wen YH, Wu MH, et al. Improving Multi-Tumor Biomarker Health Check-Up Tests with Machine Learning Algorithms. Cancers. 2020 Jun;12(6):1442. [CrossRef]
  7. Loud JT, Murphy J. Cancer Screening and Early Detection in the 21st Century. Semin Oncol Nurs. 2017 May;33(2):121–8. [CrossRef]
  8. Uncertainty Around Tests That Screen for Many Cancers - NCI [Internet]. 2022 [cited 2023 Oct 25]. Available from: https://www.cancer.gov/news-events/cancer-currents-blog/2022/finding-cancer-early-mced-tests.
  9. Huguet N, Angier H, Rdesinski R, Hoopes M, Marino M, Holderness H, et al. Cervical and colorectal cancer screening prevalence before and after Affordable Care Act Medicaid expansion. Prev Med. 2019 Jul;124:91–7. [CrossRef]
  10. Hackshaw A, Cohen SS, Reichert H, Kansal AR, Chung KC, Ofman JJ. Estimating the population health impact of a multi-cancer early detection genomic blood test to complement existing screening in the US and UK. Br J Cancer. 2021 Nov;125(10):1432–42. [CrossRef]
  11. Wang HY, Chang SC, Lin WY, Chen CH, Chiang SH, Huang KY, et al. Machine Learning-Based Method for Obesity Risk Evaluation Using Single-Nucleotide Polymorphisms Derived from Next-Generation Sequencing. J Comput Biol. 2018 Dec;25(12):1347–60. [CrossRef]
  12. Tseng YJ, Huang CE, Wen CN, Lai PY, Wu MH, Sun YC, et al. Predicting breast cancer metastasis by using serum biomarkers and clinicopathological data with machine learning technologies. Int J Med Inf [Internet]. 2019; Available from: https://www.sciencedirect.com/science/article/abs/pii/S1386505618311213?via%3Dihub. [CrossRef]
  13. Tseng YJ, Wang HY, Lin TW, Lu JJ, Hsieh CH, Liao CT. Development of a Machine Learning Model for Survival Risk Stratification of Patients With Advanced Oral Cancer. JAMA Netw Open. 2020 Aug 21;3(8):e2011768. [CrossRef]
  14. Voitechovič E, Pauliukaite R. Electrochemical multisensor systems and arrays in the era of artificial intelligence. Curr Opin Electrochem. 2023 Oct 13;101411. [CrossRef]
  15. Cohen JD, Li L, Wang Y, Thoburn C, Afsari B, Danilova L, et al. Detection and localization of surgically resectable cancers with a multi-analyte blood test. Science. 2018 Feb 23;359(6378):926–30. [CrossRef]
  16. Wang HY, Hsieh CH, Wen CN, Wen YH, Chen CH, Lu JJ. Cancers Screening in an Asymptomatic Population by Using Multiple Tumour Markers. PLoS One. 2016;11(6):e0158285. [CrossRef]
  17. Luan Y, Zhong G, Li S, Wu W, Liu X, Zhu D, et al. A panel of seven protein tumour markers for effective and affordable multi-cancer early detection by artificial intelligence: a large-scale and multicentre case–control study. eClinicalMedicine [Internet]. 2023 Jul 1 [cited 2023 Nov 6];61. Available from: https://www.thelancet.com/journals/eclinm/article/PIIS2589-5370(23)00218-3/fulltext. [CrossRef]
  18. Schrag D, Beer TM, McDonnell CH, Nadauld L, Dilaveri CA, Reid R, et al. Blood-based tests for multicancer early detection (PATHFINDER): a prospective cohort study. The Lancet. 2023 Oct;402(10409):1251–60. [CrossRef]
  19. Ahlquist, DA. Ahlquist DA. Universal cancer screening: revolutionary, rational, and realizable. Npj Precis Oncol. 2018 Oct 29;2(1):1–5. [CrossRef]
  20. Neal RD, Johnson P, Clarke CA, Hamilton SA, Zhang N, Kumar H, et al. Cell-Free DNA–Based Multi-Cancer Early Detection Test in an Asymptomatic Screening Population (NHS-Galleri): Design of a Pragmatic, Prospective Randomised Controlled Trial. Cancers. 2022 Jan;14(19):4818.
  21. Hall, IJ. Hall IJ. Patterns and Trends in Cancer Screening in the United States. Prev Chronic Dis [Internet]. 2018 [cited 2023 Oct 25];15. Available from: https://www.cdc.gov/pcd/issues/2018/17_0465.htm. [CrossRef]
  22. Zugni F, Padhani AR, Koh DM, Summers PE, Bellomi M, Petralia G. Whole-body magnetic resonance imaging (WB-MRI) for cancer screening in asymptomatic subjects of the general population: review and recommendations. Cancer Imaging. 2020 May 11;20(1):34. [CrossRef]
  23. Brenner DJ, Elliston CD. Estimated Radiation Risks Potentially Associated with Full-Body CT Screening. Radiology. 2004 Sep;232(3):735–8. [CrossRef]
  24. Brito-Rocha T, Constâncio V, Henrique R, Jerónimo C. Shifting the Cancer Screening Paradigm: The Rising Potential of Blood-Based Multi-Cancer Early Detection Tests. Cells. 2023 Jan;12(6):935. [CrossRef]
  25. Furtado CD, Aguirre DA, Sirlin CB, Dang D, Stamato SK, Lee P, et al. Whole-Body CT Screening: Spectrum of Findings and Recommendations in 1192 Patients. Radiology. 2005 Nov;237(2):385–94. [CrossRef]
  26. Schöder H, Gönen M. Screening for cancer with PET and PET/CT: potential and limitations. J Nucl Med Off Publ Soc Nucl Med. 2007 Jan;48 Suppl 1:4S-18S.
  27. Han W, Kong R, Wang N, Bao W, Mao X, Lu J. Confocal Laser Endomicroscopy for Detection of Early Upper Gastrointestinal Cancer. Cancers. 2023 Jan;15(3):776. [CrossRef]
  28. Kim SY, Kim HS, Park HJ. Adverse events related to colonoscopy: Global trends and future challenges. World J Gastroenterol. 2019 Jan 14;25(2):190–204. [CrossRef]
  29. Barbany G, Arthur C, Liedén A, Nordenskjöld M, Rosenquist R, Tesi B, et al. Cell-free tumour DNA testing for early detection of cancer – a potential future tool. J Intern Med. 2019;286(2):118–36. [CrossRef]
  30. Bettegowda C, Sausen M, Leary RJ, Kinde I, Wang Y, Agrawal N, et al. Detection of Circulating Tumor DNA in Early- and Late-Stage Human Malignancies. Sci Transl Med. 2014 Feb 19;6(224):224ra24-224ra24.
  31. Cree IA, Uttley L, Buckley Woods H, Kikuchi H, Reiman A, Harnan S, et al. The evidence base for circulating tumour DNA blood-based biomarkers for the early detection of cancer: a systematic mapping review. BMC Cancer. 2017 Oct 23;17(1):697. [CrossRef]
  32. Aravanis AM, Lee M, Klausner RD. Next-Generation Sequencing of Circulating Tumor DNA for Early Cancer Detection. Cell. 2017 Feb 9;168(4):571–4. [CrossRef]
  33. Wen YH, Chang PY, Hsu CM, Wang HY, Chiu CT, Lu JJ. Cancer screening through a multi-analyte serum biomarker panel during health check-up examinations: Results from a 12-year experience. Clin Chim Acta Int J Clin Chem. 2015 Oct 23;450:273–6. [CrossRef]
  34. Molina R, Marrades RM, Augé JM, Escudero JM, Viñolas N, Reguart N, et al. Assessment of a Combined Panel of Six Serum Tumor Markers for Lung Cancer. Am J Respir Crit Care Med. 2016 Feb 15;193(4):427–37. [CrossRef]
  35. Wu X, Wang HY, Shi P, Sun R, Wang X, Luo Z, et al. Long short-term memory model - A deep learning approach for medical data with irregularity in cancer predication with tumor markers. Comput Biol Med. 2022 May;144:105362. [CrossRef]
  36. Bodaghi A, Fattahi N, Ramazani A. Biomarkers: Promising and valuable tools towards diagnosis, prognosis and treatment of Covid-19 and other diseases. Heliyon. 2023 Feb;9(2):e13323. [CrossRef]
  37. Hartl J, Kurth F, Kappert K, Horst D, Mülleder M, Hartmann G, et al. Quantitative protein biomarker panels: a path to improved clinical practice through proteomics. EMBO Mol Med. 2023 Apr 11;15(4):e16061. [CrossRef]
  38. Messner CB, Demichev V, Wendisch D, Michalick L, White M, Freiwald A, et al. Ultra-High-Throughput Clinical Proteomics Reveals Classifiers of COVID-19 Infection. Cell Syst. 2020 Jul 22;11(1):11-24.e4. [CrossRef]
  39. Landegren U, Hammond M. Cancer diagnostics based on plasma protein biomarkers: hard times but great expectations. Mol Oncol. 2021 Jun;15(6):1715–26. [CrossRef]
  40. Vignoli A, Tenori L, Morsiani C, Turano P, Capri M, Luchinat C. Serum or Plasma (and Which Plasma), That Is the Question. J Proteome Res. 2022 Apr 1;21(4):1061–72.
  41. Rai AJ, Gelfand CA, Haywood BC, Warunek DJ, Yi J, Schuchard MD, et al. HUPO Plasma Proteome Project specimen collection and handling: towards the standardization of parameters for plasma proteome samples. Proteomics. 2005 Aug;5(13):3262–77.
  42. Wong YL, Ramanathan A, Yuen KM, Mustafa WMW, Abraham MT, Tay KK, et al. Comparative sera proteomics analysis of differentially expressed proteins in oral squamous cell carcinoma. PeerJ. 2021;9:e11548. [CrossRef]
  43. Bader JM, Albrecht V, Mann M. MS-Based Proteomics of Body Fluids: The End of the Beginning. Mol Cell Proteomics MCP. 2023 Jul;22(7):100577. [CrossRef]
  44. Fu Q, Kowalski MP, Mastali M, Parker SJ, Sobhani K, van den Broek I, et al. Highly Reproducible Automated Proteomics Sample Preparation Workflow for Quantitative Mass Spectrometry. J Proteome Res. 2018 Jan 5;17(1):420–8. [CrossRef]
  45. Wang Z, Tober-Lau P, Farztdinov V, Lemke O, Schwecke T, Steinbrecher S, et al. The human host response to monkeypox infection: a proteomic case series study. EMBO Mol Med. 2022 Nov 8;14(11):e16643. [CrossRef]
  46. Percy AJ, Yang J, Chambers AG, Mohammed Y, Miliotis T, Borchers CH. Protocol for Standardizing High-to-Moderate Abundance Protein Biomarker Assessments Through an MRM-with-Standard-Peptides Quantitative Approach. Adv Exp Med Biol. 2016;919:515–30.
  47. Füzéry AK, Levin J, Chan MM, Chan DW. Translation of proteomic biomarkers into FDA approved cancer diagnostics: issues and challenges. Clin Proteomics. 2013 Oct 2;10(1):13. [CrossRef]
  48. Van Gorp T, Cadron I, Despierre E, Daemen A, Leunen K, Amant F, et al. HE4 and CA125 as a diagnostic test in ovarian cancer: prospective validation of the Risk of Ovarian Malignancy Algorithm. Br J Cancer. 2011 Mar 1;104(5):863–70. [CrossRef]
  49. Kim YS, Kang KN, Shin YS, Lee JE, Jang JY, Kim CW. Diagnostic value of combining tumor and inflammatory biomarkers in detecting common cancers in Korea. Clin Chim Acta. 2021 May 1;516:169–78. [CrossRef]
  50. Salvi S, Gurioli G, De Giorgi U, Conteduca V, Tedaldi G, Calistri D, et al. Cell-free DNA as a diagnostic marker for cancer: current insights. OncoTargets Ther. 2016;9:6549–59. [CrossRef]
  51. De Mattos-Arruda L, Caldas C. Cell-free circulating tumour DNA as a liquid biopsy in breast cancer. Mol Oncol. 2016 Mar;10(3):464–74. [CrossRef]
  52. Gao Q, Zeng Q, Wang Z, Li C, Xu Y, Cui P, et al. Circulating cell-free DNA for cancer early detection. The Innovation. 2022 May 6;3(4):100259. [CrossRef]
  53. Bronkhorst AJ, Ungerer V, Holdenrieder S. The emerging role of cell-free DNA as a molecular marker for cancer management. Biomol Detect Quantif. 2019 Mar;17:100087. [CrossRef]
  54. Kalendar R, Shustov AV, Akhmetollayev I, Kairov U. Designing Allele-Specific Competitive-Extension PCR-Based Assays for High-Throughput Genotyping and Gene Characterization. Front Mol Biosci. 2022;9:773956. [CrossRef]
  55. Ahmad E, Ali A, Nimisha null, Kumar Sharma A, Ahmed F, Mehdi Dar G, et al. Molecular approaches in cancer. Clin Chim Acta Int J Clin Chem. 2022 Dec 1;537:60–73. [CrossRef]
  56. Ito K, Suzuki Y, Saiki H, Sakaguchi T, Hayashi K, Nishii Y, et al. Utility of Liquid Biopsy by Improved PNA-LNA PCR Clamp Method for Detecting EGFR Mutation at Initial Diagnosis of Non-Small-Cell Lung Cancer: Observational Study of 190 Consecutive Cases in Clinical Practice. Clin Lung Cancer. 2018 Mar;19(2):181–90. [CrossRef]
  57. Heitzer E, Haque IS, Roberts CES, Speicher MR. Current and future perspectives of liquid biopsies in genomics-driven oncology. Nat Rev Genet. 2019 Feb;20(2):71–88. [CrossRef]
  58. Zhai J, Arikit S, Simon SA, Kingham BF, Meyers BC. Rapid construction of parallel analysis of RNA end (PARE) libraries for Illumina sequencing. Methods San Diego Calif. 2014 May 1;67(1):84–90. [CrossRef]
  59. Belic J, Koch M, Ulz P, Auer M, Gerhalter T, Mohan S, et al. Rapid Identification of Plasma DNA Samples with Increased ctDNA Levels by a Modified FAST-SeqS Approach. Clin Chem. 2015 Jun;61(6):838–49. [CrossRef]
  60. Murtaza M, Dawson SJ, Tsui DWY, Gale D, Forshew T, Piskorz AM, et al. Non-invasive analysis of acquired resistance to cancer therapy by sequencing of plasma DNA. Nature. 2013 May 2;497(7447):108–12. [CrossRef]
  61. Newman AM, Lovejoy AF, Klass DM, Kurtz DM, Chabon JJ, Scherer F, et al. Integrated digital error suppression for improved detection of circulating tumor DNA. Nat Biotechnol. 2016 May;34(5):547–55. [CrossRef]
  62. Lanman RB, Mortimer SA, Zill OA, Sebisanovic D, Lopez R, Blau S, et al. Analytical and Clinical Validation of a Digital Sequencing Panel for Quantitative, Highly Accurate Evaluation of Cell-Free Circulating Tumor DNA. PloS One. 2015;10(10):e0140712. [CrossRef]
  63. García-Foncillas J, Alba E, Aranda E, Díaz-Rubio E, López-López R, Tabernero J, et al. Incorporating BEAMing technology as a liquid biopsy into clinical practice for the management of colorectal cancer patients: an expert taskforce review. Ann Oncol Off J Eur Soc Med Oncol. 2017 Dec 1;28(12):2943–9. [CrossRef]
  64. Fiala C, Diamandis EP. Utility of circulating tumor DNA in cancer diagnostics with emphasis on early detection. BMC Med. 2018 Oct 2;16(1):166. [CrossRef]
  65. Manokhina I, Singh TK, Peñaherrera MS, Robinson WP. Quantification of cell-free DNA in normal and complicated pregnancies: overcoming biological and technical issues. PloS One. 2014;9(7):e101500. [CrossRef]
  66. Davies MPA, Sato T, Ashoor H, Hou L, Liloglou T, Yang R, et al. Plasma protein biomarkers for early prediction of lung cancer. eBioMedicine. 2023 Jul 1;93:104686. [CrossRef]
  67. Trinidad CV, Pathak HB, Cheng S, Tzeng SC, Madan R, Sardiu ME, et al. Lineage specific extracellular vesicle-associated protein biomarkers for the early detection of high grade serous ovarian cancer. Sci Rep. 2023 Oct 26;13(1):18341. [CrossRef]
  68. Tivey A, Church M, Rothwell D, Dive C, Cook N. Circulating tumour DNA — looking beyond the blood. Nat Rev Clin Oncol. 2022 Sep;19(9):600–12. [CrossRef]
  69. Han SJ, Yoo S, Choi SH, Hwang EH. Actual half-life of alpha-fetoprotein as a prognostic tool in pediatric malignant tumors. Pediatr Surg Int. 1997;12(8):599–602.
  70. Riedinger JM, Wafflart J, Ricolleau G, Eche N, Larbre H, Basuyau JP, et al. CA 125 half-life and CA 125 nadir during induction chemotherapy are independent predictors of epithelial ovarian cancer outcome: results of a French multicentric study. Ann Oncol Off J Eur Soc Med Oncol. 2006 Aug;17(8):1234–8. [CrossRef]
  71. Halner A, Hankey L, Liang Z, Pozzetti F, Szulc DA, Mi E, et al. DEcancer: Machine learning framework tailored to liquid biopsy based cancer detection and biomarker signature selection. iScience. 2023 May 19;26(5):106610. [CrossRef]
  72. Lin WY, Chen CH, Tseng YJ, Tsai YT, Chang CY, Wang HY, et al. Predicting post-stroke activities of daily living through a machine learning-based approach on initiating rehabilitation. Int J Med Inf. 2018/02/10 ed. 2018 Mar;111:159–64. [CrossRef]
  73. Uddin S, Khan A, Hossain ME, Moni MA. Comparing different supervised machine learning algorithms for disease prediction. BMC Med Inform Decis Mak. 2019 Dec 21;19(1):281. [CrossRef]
  74. Cruz JA, Wishart DS. Applications of machine learning in cancer prediction and prognosis. Cancer Inform. 2007 Feb 11;2:59–77. [CrossRef]
  75. Wang H, Huang G. Application of support vector machine in cancer diagnosis. Med Oncol Northwood Lond Engl. 2011 Dec;28 Suppl 1:S613-618. [CrossRef]
  76. Wang HY, Chung CR, Chen CJ, Lu KP, Tseng YJ, Chang TH, et al. Clinically Applicable System for Rapidly Predicting Enterococcus faecium Susceptibility to Vancomycin. Microbiol Spectr. 2021 Dec 22;9(3):e0091321. [CrossRef]
  77. Christodoulou E, Ma J, Collins GS, Steyerberg EW, Verbakel JY, Van Calster B. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol. 2019/02/15 ed. 2019 Jun;110:12–22. [CrossRef]
  78. Liu MC, Oxnard GR, Klein EA, Swanton C, Seiden MV, CCGA Consortium. Sensitive and specific multi-cancer detection and localization using methylation signatures in cell-free DNA. Ann Oncol Off J Eur Soc Med Oncol. 2020 Jun;31(6):745–59. [CrossRef]
  79. Cebekhulu E, Onumanyi AJ, Isaac SJ. Performance Analysis of Machine Learning Algorithms for Energy Demand–Supply Prediction in Smart Grids. Sustainability. 2022 Jan;14(5):2546. [CrossRef]
  80. Yu JR, Chen CH, Huang TW, Lu JJ, Chung CR, Lin TW, et al. Energy Efficiency of Inference Algorithms for Clinical Laboratory Data Sets: Green Artificial Intelligence Study. J Med Internet Res. 2022 Jan 25;24(1):e28036.
  81. Kelly CJ, Karthikesalingam A, Suleyman M, Corrado G, King D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 2019 Oct 29;17(1):195. [CrossRef]
  82. Liu X, Faes L, Kale AU, Wagner SK, Fu DJ, Bruynseels A, et al. A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health. 2019 Oct;1(6):e271–97. [CrossRef]
  83. Yang S, Zhu F, Ling X, Liu Q, Zhao P. Intelligent Health Care: Applications of Deep Learning in Computational Medicine. Front Genet. 2021;12:607471. [CrossRef]
  84. Wan KW, Wong CH, Ip HF, Fan D, Yuen PL, Fong HY, et al. Evaluation of the performance of traditional machine learning algorithms, convolutional neural network and AutoML Vision in ultrasound breast lesions classification: a comparative study. Quant Imaging Med Surg. 2021 Apr;11(4):1381–93. [CrossRef]
  85. Rodrigues AJ, Schonfeld E, Varshneya K, Stienen MN, Staartjes VE, Jin MC, et al. Comparison of Deep Learning and Classical Machine Learning Algorithms to Predict Postoperative Outcomes for Anterior Cervical Discectomy and Fusion Procedures With State-of-the-art Performance. Spine. 2022 Dec 1;47(23):1637–44. [CrossRef]
  86. Chung CR, Wang HY, Lien F, Tseng YJ, Chen CH, Lee TY, et al. Incorporating Statistical Test and Machine Intelligence Into Strain Typing of Staphylococcus haemolyticus Based on Matrix-Assisted Laser Desorption Ionization-Time of Flight Mass Spectrometry. Front Microbiol. 2019/10/02 ed. 2019;10:2120. [CrossRef]
  87. Feng J, Phillips RV, Malenica I, Bishara A, Hubbard AE, Celi LA, et al. Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare. NPJ Digit Med. 2022 May 31;5:66. [CrossRef]
  88. Editor, MB. Editor MB. What are T Values and P Values in Statistics? [Internet]. [cited 2023 Dec 13]. Available from: https://blog.minitab.com/en/statistics-and-quality-data-analysis/what-are-t-values-and-p-values-in-statistics.
  89. Parsons, VL. Parsons VL. Stratified Sampling. In: Wiley StatsRef: Statistics Reference Online [Internet]. John Wiley & Sons, Ltd; 2017 [cited 2023 Dec 29]. p. 1–11. Available from: https://onlinelibrary.wiley.com/doi/abs/10.1002/9781118445112.stat05999.pub2.
  90. Eche T, Schwartz LH, Mokrane FZ, Dercle L. Toward Generalizability in the Deployment of Artificial Intelligence in Radiology: Role of Computation Stress Testing to Overcome Underspecification. Radiol Artif Intell. 2021 Oct 27;3(6):e210097. [CrossRef]
  91. Lapić I, Šegulja D, Dukić K, Bogić A, Lončar Vrančić A, Komljenović S, et al. Analytical validation of 39 clinical chemistry tests and 17 immunoassays on the Alinity analytical system. Scand J Clin Lab Invest. 2022 May;82(3):199–209. [CrossRef]
  92. Sun Y, Li J, Xu Y, Zhang T, Wang X. Deep learning versus conventional methods for missing data imputation: A review and comparative study. Expert Syst Appl. 2023 Oct 1;227:120201. [CrossRef]
  93. Ndugga N, Published SA. Disparities in Health and Health Care: 5 Key Questions and Answers [Internet]. KFF. 2023 [cited 2023 Dec 24]. Available from: https://www.kff.org/racial-equity-and-health-policy/issue-brief/disparities-in-health-and-health-care-5-key-question-and-answers/.
  94. Kruk ME, Gage AD, Arsenault C, Jordan K, Leslie HH, Roder-DeWan S, et al. High-quality health systems in the Sustainable Development Goals era: time for a revolution. Lancet Glob Health. 2018;6:e1196–252. [CrossRef]
  95. Siegel RL, Miller KD, Wagle NS, Jemal A. Cancer statistics, 2023. CA Cancer J Clin. 2023 Jan;73(1):17–48.
  96. Youssef A, Ng MY, Long J, Hernandez-Boussard T, Shah N, Miner A, et al. Organizational Factors in Clinical Data Sharing for Artificial Intelligence in Health Care. JAMA Netw Open. 2023 Dec 19;6(12):e2348422. [CrossRef]
  97. Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023 Jan 3;10(1):1.
  98. Wang, H. Wang H. DARTA - A Permissionless Biomarker Data Marketplace [Internet]. 2023 [cited 2023 Dec 24]. Available from: https://github.com/HsinYaoWang/DARTA.
  99. Shah NH, Halamka JD, Saria S, Pencina M, Tazbaz T, Tripathi M, et al. A Nationwide Network of Health AI Assurance Laboratories. JAMA [Internet]. 2023 Dec 20 [cited 2023 Dec 27]; Available from: . [CrossRef]
  100. Gregg AR, Skotko BG, Benkendorf JL, Monaghan KG, Bajaj K, Best RG, et al. Noninvasive prenatal screening for fetal aneuploidy, 2016 update: a position statement of the American College of Medical Genetics and Genomics. Genet Med. 2016 Oct 1;18(10):1056–65.
Figure 1. Probability distribution difference between cancer cases and normal (healthy) cases in A. case-control study dataset, and B. real-world cancer screening dataset. In a case-control study dataset, cancer cases and non-cancer cases are well-defined at the time of enrollment. The risk score distributions of cancer and non-cancer cases would be apparently different. In this case, any cutoff value in between Cutoff.1 and Cutoff.2 is fine to have a perfect predictive performance. By contrast, the risk score distributions of cancer and non-cancer cases in real-world cancer screening dataset overlap more and the optimal diagnostic cutoff is much more narrow than those in case-control study. The illustrative plots demonstrate the reason why MCED AI models that are trained by using data of case-control studies would have suboptimal predictive performance in a real-world cancer screening.
Figure 1. Probability distribution difference between cancer cases and normal (healthy) cases in A. case-control study dataset, and B. real-world cancer screening dataset. In a case-control study dataset, cancer cases and non-cancer cases are well-defined at the time of enrollment. The risk score distributions of cancer and non-cancer cases would be apparently different. In this case, any cutoff value in between Cutoff.1 and Cutoff.2 is fine to have a perfect predictive performance. By contrast, the risk score distributions of cancer and non-cancer cases in real-world cancer screening dataset overlap more and the optimal diagnostic cutoff is much more narrow than those in case-control study. The illustrative plots demonstrate the reason why MCED AI models that are trained by using data of case-control studies would have suboptimal predictive performance in a real-world cancer screening.
Preprints 95104 g001
Figure 2. Special considerations on developing MCED AI models. A. Paucity of cancer cases in the real-world cancer screening scenario. In a real-world cancer screening dataset, the ratio of cancer cases versus non-cancer cases is typically around 1:100. The ratio is extremely unbalanced for training an AI model. Over-sampling of cancer cases or under-sampling of non-cancer cases is commonly used data processing method to create a balanced dataset for AI models training. B. Data processing for training, validating, and independently testing MCED AI models. In under-sampling strategy, stratified sampling can be used to create a balanced training dataset. By contrast, the cancer versus non-cancer cases ratio is good to be kept the same as the original dataset for both validation and independent testing in order to have accurate estimation of diagnostic metrics. Mc: cancer cases; mc: cancer cases in 5-fold split datasets; Mb: non-cancer cases; mb: non-cancer cases in 5-fold split datasets that are sampled from Mb by using stratified random sampling; Nc and Nb: cancer cases and non-cancer cases in an independent dataset. C. Different approaches for independent testing of MCED AI models. Location-wise independent testing can be used to test the generalizability of an AI model across different locations. Time-wise independent testing can be used to recurrently test the robustness of an AI model in different periods of time.
Figure 2. Special considerations on developing MCED AI models. A. Paucity of cancer cases in the real-world cancer screening scenario. In a real-world cancer screening dataset, the ratio of cancer cases versus non-cancer cases is typically around 1:100. The ratio is extremely unbalanced for training an AI model. Over-sampling of cancer cases or under-sampling of non-cancer cases is commonly used data processing method to create a balanced dataset for AI models training. B. Data processing for training, validating, and independently testing MCED AI models. In under-sampling strategy, stratified sampling can be used to create a balanced training dataset. By contrast, the cancer versus non-cancer cases ratio is good to be kept the same as the original dataset for both validation and independent testing in order to have accurate estimation of diagnostic metrics. Mc: cancer cases; mc: cancer cases in 5-fold split datasets; Mb: non-cancer cases; mb: non-cancer cases in 5-fold split datasets that are sampled from Mb by using stratified random sampling; Nc and Nb: cancer cases and non-cancer cases in an independent dataset. C. Different approaches for independent testing of MCED AI models. Location-wise independent testing can be used to test the generalizability of an AI model across different locations. Time-wise independent testing can be used to recurrently test the robustness of an AI model in different periods of time.
Preprints 95104 g002aPreprints 95104 g002b
Figure 3. Challenges for implementation of MCED AIs. Beside cancer early detection, a MCED AI product should also provide a lot of clinically-relevant information in order to successfully integrate MCED AIs into current clinical workflow of diagnosing and treating cancers.
Figure 3. Challenges for implementation of MCED AIs. Beside cancer early detection, a MCED AI product should also provide a lot of clinically-relevant information in order to successfully integrate MCED AIs into current clinical workflow of diagnosing and treating cancers.
Preprints 95104 g003
Table 1. Serum biomarker-based MCED AI products on the market. RWD: real world dataset; CCD: case-control dataset; TOO: tissue of origin.
Table 1. Serum biomarker-based MCED AI products on the market. RWD: real world dataset; CCD: case-control dataset; TOO: tissue of origin.
MCED Products Biomarkers Algorithms Model development Report
Gallery (18,78) cfDNA methylation Logistic regression Train: CCD
Validation:
  • CCD based independent testing
  • Prospective RWD
  • cancer detected/not detected
  • TOO
OneTest (6,35) Protein biomarkers Classical ML algorithmsLong short term memory algorithm Train: RWD
Validation:
  • RWD based cross-validation
  • RWD location-wise independent testing
  • cancer detected/not detected
  • categorized risk
  • TOO
OncoSeek (17) Protein biomarkers Classical ML algorithms Train: CCD
Validation:
  • CCD based cross-validation
  • CCD based independent testing
  • cancer detected/not detected
  • TOO
CancerSeek (15) cfDNA + protein biomarkers Logistic regression Train: CCD
Validation:
  • CCD based cross-validation
  • cancer detected/not detected
  • TOO
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.
Prerpints.org logo

Preprints.org is a free preprint server supported by MDPI in Basel, Switzerland.

Subscribe

© 2024 MDPI (Basel, Switzerland) unless otherwise stated