A Computational Readiness Audit of Digital Patient Reviews

1. Introduction

Digital Patient computing is often presented as a sequence: acquire data, construct a model, produce an output. Real systems are less tidy. A construct defines the subject being represented. A model makes a task-specific approximation. A platform supplies data services, interfaces, software, and operational support. A realization configures these pieces for one intended use. The Digital Patient literature has long treated modeling, visualization, systems biology, physics-based simulation, statistical reasoning, and software as complementary capabilities rather than as one canonical implementation [1]. For this review, a pipeline is therefore an analytic frame [2,3,4,5].

Three distinctions determine whether a pipeline claim is technically meaningful. First, simulating a time-dependent process is different from updating the model with newly acquired patient data. Second, sending observations into software is different from assimilating them into a state estimate or parameter set. Third, displaying an output to a clinician is different from clinician-mediated feedback, and both differ from automatic closed-loop actuation [6] Each transition introduces interfaces, assumptions, latency, failure modes, and evidence requirements.

Credibility guidance for computational modeling links the amount and type of evidence to context of use, model influence, decision consequence, and resulting model risk [7,8,9,10]. Digital-twin guidance adds interaction, updating, and lifecycle behavior to that assurance problem [11]. We used those principles to audit two registered questions: what computational elements are reported, and how much directly supportable evidence connects those elements to a credible system?

2. Methods

2.1. Protocol, Registration, and Reporting Framework

This paper is a focused analysis of a shared systematic overview registered in PROSPERO (CRD420261458411). It did not conduct a separate search, create an independent cohort, or make an additional registration. The paper-specific plan assigns specific non-overlapping research questions to computational construction and to validation and fidelity; intended functions are used only as output context following PRISMA 2020 guideline [12,13]. The focused plan was recorded after assembly of the evidence map and before submission as a transparent reporting amendment. While companion reports evaluate conceptual terminology and clinical adoption governance, this manuscript focuses exclusively on the technical construction, multimodal data integration, model coupling, interoperability layers, and Verification, Validation, and Uncertainty Quantification (VVUQ) across the computational pipeline.

The parent protocol included English-language reviews published from 2020 onward that examined patient-linked physiological, biomechanical, biological, or adjacent computational representations. Web of Science, PubMed, IEEE Xplore, and Google Scholar were searched, with reference-list and designated preprint-server supplementation. The search was conducted in October 2025; no later search was incorporated. Of 247 reports assessed for eligibility, 183 were excluded and 64 entered extraction (Figure 1). A bibliographic audit then removed three duplicate reports, yielding 61 unique reviews.

2.2. Data Collection

CADIMA, a web-based evidence-synthesis platform, was used to manage and document the systematic-review process (Kohl et al., 2018) [14]. Data were collected on data sources, model types, coupling, temporal orientation, modeling approaches, integration standards, computing infrastructure, tools or platforms, validation methods and data, fidelity metrics, reproducibility, source-review appraisal, intended functions, maturity wording, and limitations. Candidate categories were generated by prespecified phrase matching. A review contributed no more than one count to a category, but categories were non-exclusive.

Inputs were grouped as imaging; sensors, wearables, or internet of things (IoT); electronic health records (EHR) or clinical records; omics; laboratory or biomarkers; lifestyle or environment; and simulation or literature. Model families were AI or machine learning, mechanistic or physics-based, hybrid, simulation, and statistical or probabilistic. Architecture fields included cloud, edge, hybrid cloud–edge, high-performance, and federated or distributed computing. Interface fields included DICOM, HL7, FHIR, OMOP, and named ontologies or identifiers.

2.3. Data Processing and Semantic Audit

We reproduced the scripted classifications and manually verified each row-level match against the full-text articles. Manual verification constituted the primary relabeling step, with scripted classifications serving as an initial screening aid. Component mentions confirmed during this process were retained as review-level observations. Four families of mechanically reproducible totals were stopped: validation, reproducibility, prospective evidence, and maturity. Their rules treated negated claims, proposed future work, diagnostic sensitivity, reproducibility of review methods, forecasting language, and adjacent technologies as affirmative system evidence. Reporting those totals would have converted a reproducible error into a precise-looking result. This stop rule is central to the analysis. Missing or ambiguous evidence was not reclassified as failure, and an unreliable aggregate was not replaced by undocumented judgment. Instead, validation was examined through field completeness, review-appraisal coverage, and interpretable within-review findings. Complete record-level recoding remains necessary before cohort frequencies can be claimed.

2.4. Synthesis Methods and Interpretation Constraints

The pipeline map displays marginal frequencies. Its arrows organize categories; they do not assert causal flow, mass balance, integration, or even co-occurrence in one primary system. Temporal and directional wording was used only as context. Virtual cohorts were recognized as computational evidence but not as evidence of longitudinal connection to one index patient.

Two appraisals were kept separate. A source review could formally appraise its included studies. Independently, the meta-review applied AMSTAR 2 or SANRA where appropriate, with one qualitative-only assessment [15,16]. Neither measure evaluates a model’s numerical correctness or clinical fitness.

3. Results

3.1. Layer I—Inputs, Inference Engines, and Intended Outputs

The final evidence base comprised 61 unique reviews after three post-selection duplicates were removed from the 64 records selected for extraction (Table 1). Publication increased markedly over the study period, with most reviews published in 2024 (n = 19) or 2025 (n = 27), compared with five in 2022, eight in 2023, and two in 2026. The reviews covered a broad range of clinical domains, most frequently cardiovascular applications (n = 32), oncology (n = 25), surgery, orthopedics, and dentistry (n = 24), neurology and neuroscience (n = 18), and diabetes and endocrine applications (n = 17). Instrument-based methodological appraisal was available for 39 reviews, whereas one had qualitative-only appraisal and 21 lacked an available appraisal. These appraisal results characterize the methodological context of the review evidence base and should not be interpreted as assessments of model validity or fitness for purpose. Imaging was reported in 49 reviews (80.3%), sensors, wearables, or IoT in 43 (70.5%), EHR or clinical records in 29 (47.5%), omics or genomics in 24 (39.3%), lifestyle or environment in 16 (26.2%), laboratory or biomarkers in 15 (24.6%), and simulation or literature inputs in 14 (23.0%) as summarized in Table 2. AI or machine learning appeared in 50 reviews (82.0%), mechanistic or physics-based modeling in 38 (62.3%), and hybrid modeling in 33 (54.1%). Simulation appeared in 13 and statistical or probabilistic methods in nine. These frequencies support a plural computational ecosystem. They do not show that hybrid models were integrated well or that an AI and a mechanistic component exchanged valid information in one realization.

The stated output space was equally broad: treatment or therapy planning appeared in 56 reviews (91.8%), prediction, prognosis, or risk in 49 (80.3%), diagnosis or detection in 46 (75.4%), education or training in 19 (31.1%), personalized care in nine, monitoring in eight, drug development or trials in seven, and decision support in five. An output label was counted as a reported purpose, not as a demonstrated capability. Figure 2 summarizes the reported review-level frequencies for selected input, model-family, and output categories.

3.2. Layer II—Interfaces, Interoperability, and Computational Infrastructure

Cloud computing appeared in 33 reviews, hybrid cloud–edge in 17, edge in 14, high-performance computing in four, and federated or distributed computing in three. The records did not consistently structure hardware, runtime, memory, energy, end-to-end latency, availability, or approximation error. No architecture could therefore be compared on operational performance.

DICOM was named in 15 reviews, HL7 in 14, FHIR in 13, and OMOP in 11. These standards address different layers and objects. A citation to DICOM does not prove correct image-derived geometry; a FHIR endpoint does not guarantee that physiological meaning, units, identity, provenance, timing, or uncertainty survives transformation into model parameters. The rule detected no ontology or identifier category. This is a rule-dependent non-detection, not evidence that semantic resources were absent.

The unanswered interface question is end-to-end traceability. A credible report should allow a reviewer to follow a patient observation through preprocessing, segmentation or feature extraction, geometry or state definition, boundary conditions and parameters, model execution, uncertainty, output interpretation, and the action eventually supported. The review-level data showed frequent use of standards vocabulary but could not verify that chain.

3.3. Layer III—Updating, Feedback, and Control

Aggregate temporal and coupling labels were not used as implementation frequencies. A review may call a model dynamic because it solves differential equations, may call data real time because a sensor streams them, or may use “two-way” for a clinician-facing recommendation. None establishes online state assimilation or automatic control.

Focused evidence illustrates the narrower claims. Drummond et al. identified 20 dynamic systems among 80 claimed patient digital twins and nine with two-way data flow; six returned recommendations to a patient or clinician and three supported surgical navigation [17]. Tudor et al. found 18/149 systems meeting its operational-twin criteria, of which 17 relied on a human-mediated return and one autonomous cardiac system acted through a defibrillator [18]. A contextual glucose-control study (by Hovorka et al.) cited in Cappon et al. provides a more explicit control architecture: new glucose data entered an adaptive controller that calculated insulin-infusion changes [19]. These examples distinguish dynamic, two-way, human-mediated, and autonomous-control claims. They are not pooled implementation counts.

3.4. Layer IV—Validation, Reproducibility, and Assurance Evidence

Five reviews lacked validation-method text, eight lacked validation-data-source text, 13 lacked fidelity-metric text, and 20 lacked the source-review appraisal field. Even populated cells rarely distinguished training reuse, calibration, internal testing, independent external validation, or prospective evaluation.

Only 11 source reviews reported a named formal appraisal of included evidence. The meta-review separately recorded 18 AMSTAR 2 assessments, 21 SANRA assessments, one qualitative-only assessment, and 21 unassessed reviews. These appraisal counts measure review-method coverage, not model credibility.

Record-level findings revealed recurring breaks in the assurance chain. Cappon et al. reported that four of eight type 1 diabetes methods were tested only on synthetic simulator data and two implementations were open source [20]. Rodero et al. found some form of validation in 75% of 36 cardiac in silico trials but sensitivity analysis in 19%, and warned that validating a solver or component does not validate the complete study [21]. Tudor et al. found only two primary studies mentioning VVUQ [18]. Vallee et al. described predominantly internal retrospective evaluation and no prospective demonstration of improved clinical outcomes [22]. Chen et al. found open-source implementations in 18/76 drug-development articles and judged complete reproducibility generally unavailable [23]. Stefaniga et al. concluded that most oncology twinning solutions lacked adequate clinical validation (Table 3) [24].

4. Discussions

The results show that Digital Patient reviews frequently report individual inputs, model families, interoperability specifications, computing architectures, and intended outputs, but provide substantially less evidence that these components are connected, updated, validated, and governed as an operational system. Computational readiness should therefore not be inferred from the presence of a preferred technology stack or from isolated evidence of multimodal data, hybrid modeling, standards adoption, or cloud–edge deployment, but from a linked chain of claims demonstrating that each component and transition is fit for the intended decision. Table 4 presents eight assurance checkpoints spanning the definition of the decision, data fitness and traceability, update logic, implementation correctness, model identifiability, prediction credibility, transfer to practice, and lifecycle change. Because weaknesses can propagate across this chain, the depth of supporting evidence should be proportional to the context of use, model influence, decision consequence, and resulting risk [7,8,9,11,25,26].

4.1. Implications for Computational Readiness

The review literature is rich in modalities and model families. It supports the conclusion that Digital Patient computation is many-to-many and that AI and mechanistic modeling often occupy complementary roles. It does not support a prevalence claim for integrated pipelines, data assimilation, human-mediated feedback, autonomous control, or system-level credibility.

This gap is not solved by adding another component. Hybrid models, for example, can connect learned representations to physiological constraints, but every connection creates a new interface, calibration dependency, and uncertainty pathway. A high-fidelity multiphysics model may be appropriate for offline planning but impractical for continuous operation. A surrogate may meet latency requirements but needs bounded approximation error for the intended decision. Technology choice follows the question of interest; credibility follows the evidence case.

4.2. Implications for Reporting and Practice

Authors should distinguish the following terms explicitly: training, calibration, internal validation, external validation, and prospective evaluation; dynamic simulation, episodic update, and online assimilation; one-way data acquisition, clinician-facing advice, clinician-mediated feedback, and automated actuation. They should identify the patient-data partition used at each stage, carry provenance and units across transformations, report numerical and software verification, quantify sensitivity and uncertainty, and document versions and change controls.

An executed left-ventricular-assist-device VVUQ plan shows how risk, verification, sensitivity, uncertainty, and bench validation can be traced without suggesting that this level of evidence was common in the cohort [26]. Good machine-learning-practice principles add representative data, independent testing, human–AI evaluation, and lifecycle monitoring [27].

4.3. Strengths and Limitations

Strengths include a stable post-deduplication cohort, reproducible field-specific rules, row-level semantic auditing, direct checks against informative full texts, and explicit stop rules when prespecified rules failed. The analysis does not convert missing information into failure and does not convert review-level co-mention into system architecture.

The analysis reported in the current study should be interpreted relative to multiple limitations. Reviews, not implementations, were the units of analysis, and underlying-study overlap was not quantified. Free-text fields may undercount standards, computing parameters, updating, control, and evidence. Reporting in the full-text articles may have underrepresented standards, computing parameters, updating, control, and supporting evidence. State estimation, data assimilation, verification, calibration, model influence, decision consequences, external validation, latency, energy use, change control, and drift were also not reported consistently enough for reliable review-level classification. Full record-level recoding is required before validation, reproducibility, prospective-evidence, or maturity frequencies are defensible.

5. Conclusion

The computational foundation of Digital Patients cannot be judged from the number of modalities, models, or platforms named in a review. The assurance object is the connected pipeline: data with preserved meaning, verified execution, identifiable calibration, validation tied to the decision, uncertainty carried to the output, and lifecycle controls that respond to change. The literature demonstrates a substantial parts catalogue. Whether those parts form dependable clinical systems remains a record-level and ultimately implementation-level question.

6. Declarations

Copyrighted full texts and the source extraction workbook remain local and are not publicly redistributed. The cohort and review process originated in a broader evidence-map exercise and were reused here to avoid duplicate searching, screening, and extraction. The distinct analyses in this paper concern computational construction and credibility While companion reports evaluate conceptual terminology and clinical adoption governance [28].

7. Supplementary Appendix S1: Protocol and Search Method

7.1. Status and Timing

This note records the focused analysis as one paper arising from a broader evidence map conducted under a single PROSPERO-registered parent protocol (CRD420261458411). It was specified after the source evidence map had been assembled and before submission. It is not a prospective protocol, a new review, or an additional PROSPERO submission. The parent search, eligibility criteria, screening, extraction workbook, duplicate adjudication, and 61-review cohort were not changed for this analysis.

The source protocol draft and this pre-submission note are retained in the project records. The complete database search description and this note should be verified as deposited in the public reproducibility repository before submission. Stating when focused questions were formalized is consistent with transparent reporting of protocol changes [12,29].

7.2. Focused Objectives

The focused analysis asks:

• Which data sources, model families, standards, and computing architectures are reported for digital patient frameworks?
• Which input-model-output relationships are co-mentioned at review level?
• What directly supportable validation, uncertainty, reproducibility, and review-appraisal evidence is reported?
• Which standards-informed evidence checkpoints are relevant from data fitness through lifecycle operation?

7.3. Search Strategy

Table 5 reproduces the project search log. Counts retain its result (origin) notation and are not PRISMA flow counts. Reference lists of included reviews were also checked; no Boolean query applies.

PubMed — 4 October 2025; combined 1,159 before filters and 800 after filters; project-log count 794 (800)

("digital twin"[tiab:~1] OR "digital twins"[tiab:~1] OR "digital twinning"[tiab:~1] OR "virtual twin"[tiab:~1] OR "virtual twins"[tiab:~1] OR "virtual twinning"[tiab:~1] OR "digital replica"[tiab:~1] OR "digital replicas"[tiab:~1] OR "digital replication"[tiab:~1] OR "virtual human*"[tiab] OR "virtual patient"[tiab:~1] OR "virtual patients"[tiab:~1] OR "personalized computational model"[tiab:~1] OR "personalized computational models"[tiab:~1] OR "personalized computational modeling"[tiab:~1] OR "personalized computational modelling"[tiab:~1] OR "patient-specific computational model"[tiab:~1] OR "patient-specific computational models"[tiab:~1] OR "patient-specific computational modelling"[tiab:~1] OR "patient-specific computational modeling"[tiab:~1] OR "medical twin*" OR "digital patient*"[tiab] OR "patient avatar*"[tiab] OR "digital avatar*"[tiab] OR "virtual avatar*"[tiab] OR "health twin*"[tiab] OR "organ twin*"[tiab] OR "surrogate model*"[tiab] OR "virtual physiological human*"[tiab] OR "biomimetic twin"[tiab:~1] OR "biomimetic twins"[tiab:~1] OR "biomimetic twinning"[tiab:~1] OR "in silico patient*"[tiab]) AND (patient*[tiab] OR patients[MeSH] OR medic*[tiab] OR medicine[MeSH] OR "delivery of health care"[MeSH] OR healthcare[tiab] OR "health care"[tiab] OR health[MeSH] OR health[tiab] OR surgery[SH] OR "surgical procedures, operative"[MeSH] OR "general surgery"[MeSH] OR surg*[tiab] OR transplants[MeSH] OR transplant*[tiab] OR therapeutics[MeSH] OR therap*[tiab] OR treatment*[MeSH] OR "pharmaceutical preparations"[MeSH] OR drug*[tiab] OR pharmacology[MeSH] OR pharmacokinetics[MeSH] OR pharmaco*[tiab] OR neoplasms[MeSH] OR oncology[tiab] OR cancer[tiab] OR clinic*[tiab] OR humans[MeSH] OR human*[tiab] OR physiology[MeSH] OR physiology*[tiab] OR physiopathology[SH] OR pathophysiolog*[tiab] OR "biomechanical phenomena"[MeSH] OR biomechanic*[tiab] OR imaging[tiab] OR genome[MeSH] OR genomics[MeSH] OR genom*[tiab] OR "precision medicine"[MeSH] OR "precision medicine"[tiab] OR "precision health*"[tiab] OR "digital health"[MeSH] OR "digital health*"[tiab] OR "personalized medicine"[tiab] OR "equipment and supplies"[MeSH] OR "medical device*"[tiab] OR "preventive medicine"[MeSH] OR "preventative medicine"[tiab] OR "preventive medicine"[tiab]) AND ("systematic review"[PT] OR "systematic review"[tiab] OR "scoping review"[tiab] OR "meta-analysis"[PT] OR "meta-analysis"[tiab] OR "meta-review"[tiab] OR "review of reviews"[tiab] OR "umbrella review"[tiab] OR review[PT] OR review[tiab] OR "systematic overview"[tiab]) Filters: English; publication year 2020 onward.

Google Scholar — 18 October 2025; project-log count 102 (241)

("digital twin" OR "digital twins" OR "virtual twin" OR "virtual twins" OR "virtual patient" OR "digital patient" OR "in silico patient" OR "patient specific model" OR "virtual human") (health OR medical OR clinical OR healthcare OR patient OR therapy OR surgery OR drug) intitle:review

medRxiv — 18 October 2025; project-log count 25 (26)

("digital twin" OR "digital twins" OR "virtual twin" OR "virtual twins" OR "virtual patient" OR "digital patient" OR "in silico patient" OR "patient specific model" OR "virtual human") (model OR models OR modelling OR modeling) AND intitle:review AND source:medrxiv Limit recorded in the project log: since 2020.

arXiv — 18 October 2025; project-log count 146 (149)

("digital twin" OR "digital twins" OR "virtual twin" OR "virtual twins" OR "virtual patient" OR "digital patient" OR "in silico patient" OR "patient-specific model" OR "virtual human") (health OR medical OR clinical OR healthcare OR patient OR therapy OR surgery OR drug) (model OR models OR modeling OR modelling) intitle:review source:arxiv

Web of Science — 18 October 2025; project-log count 964 (1,495)

TS=(digital NEAR/1 twin) OR TS=(digital NEAR/1 twins) OR TS=(digital NEAR/1 twinning) OR TS=(virtual NEAR/1 twin) OR TS=(virtual NEAR/1 twins) OR TS=(virtual NEAR/1 twinning) OR TS=(digital NEAR/1 replica) OR TS=(digital NEAR/1 replicas) OR TS=(digital NEAR/1 replication) OR TS=("virtual human*") OR TS=(virtual NEAR/1 patient) OR TS=(virtual NEAR/1 patients) OR TS=(personalized NEAR/1 computational NEAR/1 model) OR TS=(personalized NEAR/1 computational NEAR/1 models) OR TS=(personalized NEAR/1 computational NEAR/1 modeling) OR TS=(personalized NEAR/1 computational EAR/1 modelling) OR TS=(patient-specific NEAR/1 computational NEAR/1 model) OR TS=(patient-specific NEAR/1 computational NEAR/1 models) OR TS=(patient-specific NEAR/1 computational NEAR/1 modeling) OR TS=(patient-specific NEAR/1 computational NEAR/1 modelling) OR TS=("medical twin*") OR TS=("digital patient*") OR TS=("patient avatar*") OR TS=("digital avatar*") OR TS=("virtual avatar*") OR TS=("health twin*") OR TS=("organ twin*") OR TS=("surrogate model*") OR TS=("virtual physiological human*") OR TS=(biomimetic NEAR/1 twin) OR TS=(biomimetic NEAR/1 twins) OR TS=(biomimetic NEAR/1 twinning) OR TS=("in silico patient*") AND TS=(patient* OR medic* OR healthcare OR "health care" OR health OR surg* OR transplant* OR therap* OR treatment* OR drug* OR pharmaco* OR oncology OR cancer OR clinic* OR human* OR physiology* OR pathophysiolog* OR biomechanic* OR imaging OR genom* OR "precision medicine" OR "precision health*" OR "digital health*" OR "personalized medicine" OR "medical device*" OR "preventive medicine" OR "preventative medicine") AND TS=("systematic review" OR "scoping review" OR "meta-analysis" OR "meta review" OR "review of reviews" OR "umbrella review" OR "systematic overview" OR review)

IEEE Xplore — 18 October 2025; project-log count 526 (603)

(("digital twin" OR "digital twins" OR "digital twinning" OR "virtual twin" OR "virtual twins" OR "virtual twinning" OR "digital replica" OR "digital replicas" OR "digital replication" OR "virtual human" OR "virtual humans" OR "virtual patient" OR "virtual patients" OR "personalized computational model" OR "personalized computational models" OR "personalized computational modeling" OR "personalized computational modelling" OR "patient-specific computational model" OR "patient-specific computational models" OR "patient-specific computational modeling" OR "patient-specific computational modelling" OR "medical twin" OR "medical twins" OR "digital patient" OR "digital patients" OR "patient avatar" OR "patient avatars" OR "digital avatar" OR "digital avatars" OR "virtual avatar" OR "virtual avatars" OR "health twin" OR "health twins" OR "organ twin" OR "organ twins" OR "surrogate model" OR "surrogate models" OR "virtual physiological human" OR "virtual physiological humans" OR "biomimetic twin" OR "biomimetic twins" OR "biomimetic twinning" OR "in silico patient" OR "in silico patients") AND (patient OR patients OR medical OR medicine OR healthcare OR "health care" OR health OR surgery OR surgical OR transplant OR transplants OR therapy OR therapies OR therapeutic OR treatment OR treatments OR drug OR drugs OR pharmacology OR pharmacologic OR pharmacokinetic OR oncology OR cancer OR clinical OR clinic OR human OR humans OR physiology OR physiological OR pathophysiology OR pathophysiologic OR biomechanics OR biomechanical OR imaging OR genome OR genomic OR genomics OR "precision medicine" OR "precision health" OR "digital health" OR "personalized medicine" OR "medical device" OR "medical devices" OR "preventive medicine" OR "preventative medicine") AND ("systematic review" OR "scoping review" OR "meta-analysis" OR "meta review" OR "review of reviews" OR "umbrella review" OR "systematic overview" OR review))

7.4. Evidence Map

The relationship between the source evidence-map methods and their use in this focused analysis is summarized below.

Table 6. Source evidence-map elements and their use in the focused analysis.

Element	Source evidence map	Focused analysis
Eligibility/search	Broad parent scope and source-specific queries above	No change; no new search or topic-based exclusion
Selection	Title/abstract and full-text screening; unique review report after duplicate adjudication	Same 61 reviews; overlapping primary studies not estimated
Unit of analysis	Unique review report after post-selection duplicate adjudication	Same 61 reviews
Extraction	CADIMA-aligned workbook covering data, model, standards, compute, validation, functions, and appraisal	Specific research questions focused on computational construction and validation/credibility; functions provide output context
Appraisal	AMSTAR 2, SANRA, and CADIMA-aligned review-level fields where applicable	Qualifies review conduct; does not establish system validity or readiness
Synthesis	Descriptive review-level counts; non-exclusive categories; no pooled effects	Component frequencies, verified credibility findings, semantic stop rules, and an eight-claim assurance proposal

Table 6. Source evidence-map elements and their use in the focused analysis.

Element	Source evidence map	Focused analysis
Eligibility/search	Broad parent scope and source-specific queries above	No change; no new search or topic-based exclusion
Selection	Title/abstract and full-text screening; unique review report after duplicate adjudication	Same 61 reviews; overlapping primary studies not estimated
Unit of analysis	Unique review report after post-selection duplicate adjudication	Same 61 reviews
Extraction	CADIMA-aligned workbook covering data, model, standards, compute, validation, functions, and appraisal	Specific research questions focused on computational construction and validation/credibility; functions provide output context
Appraisal	AMSTAR 2, SANRA, and CADIMA-aligned review-level fields where applicable	Qualifies review conduct; does not establish system validity or readiness
Synthesis	Descriptive review-level counts; non-exclusive categories; no pooled effects	Component frequencies, verified credibility findings, semantic stop rules, and an eight-claim assurance proposal

7.5. Analysis Plan

Counts use 61 unique reviews as the denominator and describe review discourse rather than patients, systems, deployments, or independent evidence events. Co-reporting does not establish co-occurrence in one system. Three prospective lexical flags were adjudicated in full text; none represented a prospective Digital Patient-system outcome evaluation. Adjudicated protocol amendments and deviations are detailed in Table 7.

7.6. Evidence-Map Provenance and Companion-Report Relationship

The findings reported in this paper are a focused analysis of the search, screening, extraction, and evidence map developed under one broader PROSPERO-registered review protocol. The present paper did not undertake a new search, screening process, or extraction. It uses the shared dataset to examine computational components, interface semantics, validation evidence, and assurance.

This manuscript is one of three focused papers arising from the same registered review rather than an independent review or an additional PROSPERO submission. Companion reports cross-reference the common provenance while distinguishing objectives, analyses, figures, and conclusions; shared findings are not presented as independent evidence [28].

8. Supplementary Appendix S2: Inconsistency-Resolution Protocol Summary

The source CADIMA protocol distinguishes coordinator checks from eligibility decisions. Conflicts that cannot change eligibility were verified by the coordinator. Conflicts affecting inclusion, exclusion, or coding were discussed by reviewers; unresolved uncertainty favors retention until fuller information were made available, and the coordinator adjudicated when consensus was not achieved. Eligibility decisions, reasons, and post-selection duplicate exclusions remained in the audit trail. The focused analysis introduced no new screening decisions. Coding disagreements were resolved against the full text and documented rule set, with manual overrides restricted to explicit negation or contrastive wording.

Table 8. Supplementary Appendix S3: Paper identifiers, citations, and review-level appraisal.

ID	Review	Stratum	Rating
P01	Aboarab et al. [30]	Moderate	3/5
P02	Adibi et al. [31]	Very low	1/5
P03	Akbarialiabad et al. [32]	Very low	1/5
P04	Banoub et al. [33]	Moderate	8/12
P05	Bartusik-Aebisher et al. [34]	Unavailable	-
P06	Cappon et al. [20]	Low	2/5
P07	Chaparro-Cárdenas et al. [35]	High	11/12
P08	Chumnanvej et al. [36]	Unavailable	-
P09	Coorey et al. [37]	High	12/12
P10	D'Orsi et al. [38]	Moderate	9/12
P11	De Oliveira El-Warrak et al. [39]	Unavailable	-
P12	Drummond et al. [17]	Very low	1/5
P13	Espinoza-Vinces et al. [40]	Moderate	4/5
P14	Faiella et al. [41]	Moderate	4/5
P15	Gu et al. [42]	High	10/12
P16	Bian et al. [43]	Very low	1/5
P17	John et al. [44]	Moderate	3/5
P18	Khan et al. [45]	Unavailable	-
P19	Naik et al. [46]	Very low	1/5
P20	Pinton et al. [47]	Very low	1/5
P21	Pradíes et al. [48]	Low	2/5
P22	Riahi et al. [49]	Very low	1/5
P23	Ringeval et al. [50]	Very low	1/5
P24	Rodero et al. [21]	Unavailable	-
P25	Seth et al. [51]	Low	2/5
P26	Shen et al. [52]	Unavailable	-
P27	Shen et al. [53]	Unavailable	-
P28	Ștefănigă et al. [24]	Very low	1/5
P29	Sun et al. [54]	Moderate	7/12
P30	Tasmurzayev et al. [3]	Unavailable	-
P31	Tudor et al. [18]	High	10/12
P32	Varrassi et al. [55]	Moderate	9/12
P33	Yuan et al. [56]	High	11/12
P34	Pascual et al. [57]	Moderate	7/12
P35	Wickramasinghe et al. [58]	Low	6/12
P36	Mahmud et al. [59]	High	11/12
P37	Pawar et al. [60]	Unavailable	-
P38	Park et al. [61]	High	12/12
P39	Kim et al. [4]	Moderate	4/5
P40	Daraio et al. [62]	Unavailable	-
P41	Alshahrani et al. [63]	Moderate	7/12
P42	Zou et al. [64]	High	12/12
P43	Ghosh et al. [65]	Unavailable	-
P44	Vallée et al. [22]	High	5/5
P45	Chen et al. [23]	Unavailable	-
P46	Xu et al. [66]	Unavailable	-
P47	Pellegrino et al. [67]	Unavailable	-
P48	Oulefki et al. [68]	High	12/12
P49	Cellina et al. [69]	Moderate	7/12
P50	Garanin et al. [70]	Moderate	7/12
P51	Sasitharasarma et al. [71]	Qualitative-only	-
P52	Rudnicka et al. [72]	Unavailable	-
P53	Bjelland et al. [73]	High	10/12
P54	Lauer-Schmaltz et al. [74]	Unavailable	-
P55	Maizi et al. [75]	Unavailable	-
P56	Salvi et al. [76]	Moderate	9/12
P57	Karaduman et al. [77]	Unavailable	-
P58	Pisirgen et al. [78]	Unavailable	-
P59	Zhang et al. [79]	Unavailable	-
P60	Narigina et al. [80]	Low	5/12
P61	Xames et al. [5]	Unavailable	-

Note. P01–P61 are stable identifiers assigned after post-selection deduplication. Strata are very low, low, moderate, high, qualitative-only, and unavailable. AMSTAR 2 labels are compact positions: 1/5 critically low, 2/5 low, 3/5 low-to-moderate, 4/5 moderate, and 5/5 high; 3/5 is a non-standard extracted hybrid label, and the fractions are not scores. SANRA retains its 0–12 scale. A dash indicates no numeric rating. The crosswalk is descriptive, not a validated equivalence or pooled score, and says nothing about model validity or fitness for purpose. P16, P34, and P43 are retained preprints under the protocol's specific grey-literature clause. P41 is a journal article with verified editorial dates. The P42 citation is the verified International Journal of Medical Informatics article (doi:10.1016/j.ijmedinf.2025.106138).

Table 8. Supplementary Appendix S3: Paper identifiers, citations, and review-level appraisal.

ID	Review	Stratum	Rating
P01	Aboarab et al. [30]	Moderate	3/5
P02	Adibi et al. [31]	Very low	1/5
P03	Akbarialiabad et al. [32]	Very low	1/5
P04	Banoub et al. [33]	Moderate	8/12
P05	Bartusik-Aebisher et al. [34]	Unavailable	-
P06	Cappon et al. [20]	Low	2/5
P07	Chaparro-Cárdenas et al. [35]	High	11/12
P08	Chumnanvej et al. [36]	Unavailable	-
P09	Coorey et al. [37]	High	12/12
P10	D'Orsi et al. [38]	Moderate	9/12
P11	De Oliveira El-Warrak et al. [39]	Unavailable	-
P12	Drummond et al. [17]	Very low	1/5
P13	Espinoza-Vinces et al. [40]	Moderate	4/5
P14	Faiella et al. [41]	Moderate	4/5
P15	Gu et al. [42]	High	10/12
P16	Bian et al. [43]	Very low	1/5
P17	John et al. [44]	Moderate	3/5
P18	Khan et al. [45]	Unavailable	-
P19	Naik et al. [46]	Very low	1/5
P20	Pinton et al. [47]	Very low	1/5
P21	Pradíes et al. [48]	Low	2/5
P22	Riahi et al. [49]	Very low	1/5
P23	Ringeval et al. [50]	Very low	1/5
P24	Rodero et al. [21]	Unavailable	-
P25	Seth et al. [51]	Low	2/5
P26	Shen et al. [52]	Unavailable	-
P27	Shen et al. [53]	Unavailable	-
P28	Ștefănigă et al. [24]	Very low	1/5
P29	Sun et al. [54]	Moderate	7/12
P30	Tasmurzayev et al. [3]	Unavailable	-
P31	Tudor et al. [18]	High	10/12
P32	Varrassi et al. [55]	Moderate	9/12
P33	Yuan et al. [56]	High	11/12
P34	Pascual et al. [57]	Moderate	7/12
P35	Wickramasinghe et al. [58]	Low	6/12
P36	Mahmud et al. [59]	High	11/12
P37	Pawar et al. [60]	Unavailable	-
P38	Park et al. [61]	High	12/12
P39	Kim et al. [4]	Moderate	4/5
P40	Daraio et al. [62]	Unavailable	-
P41	Alshahrani et al. [63]	Moderate	7/12
P42	Zou et al. [64]	High	12/12
P43	Ghosh et al. [65]	Unavailable	-
P44	Vallée et al. [22]	High	5/5
P45	Chen et al. [23]	Unavailable	-
P46	Xu et al. [66]	Unavailable	-
P47	Pellegrino et al. [67]	Unavailable	-
P48	Oulefki et al. [68]	High	12/12
P49	Cellina et al. [69]	Moderate	7/12
P50	Garanin et al. [70]	Moderate	7/12
P51	Sasitharasarma et al. [71]	Qualitative-only	-
P52	Rudnicka et al. [72]	Unavailable	-
P53	Bjelland et al. [73]	High	10/12
P54	Lauer-Schmaltz et al. [74]	Unavailable	-
P55	Maizi et al. [75]	Unavailable	-
P56	Salvi et al. [76]	Moderate	9/12
P57	Karaduman et al. [77]	Unavailable	-
P58	Pisirgen et al. [78]	Unavailable	-
P59	Zhang et al. [79]	Unavailable	-
P60	Narigina et al. [80]	Low	5/12
P61	Xames et al. [5]	Unavailable	-

Note. P01–P61 are stable identifiers assigned after post-selection deduplication. Strata are very low, low, moderate, high, qualitative-only, and unavailable. AMSTAR 2 labels are compact positions: 1/5 critically low, 2/5 low, 3/5 low-to-moderate, 4/5 moderate, and 5/5 high; 3/5 is a non-standard extracted hybrid label, and the fractions are not scores. SANRA retains its 0–12 scale. A dash indicates no numeric rating. The crosswalk is descriptive, not a validated equivalence or pooled score, and says nothing about model validity or fitness for purpose. P16, P34, and P43 are retained preprints under the protocol's specific grey-literature clause. P41 is a journal article with verified editorial dates. The P42 citation is the verified International Journal of Medical Informatics article (doi:10.1016/j.ijmedinf.2025.106138).