Quantum Machine Learning for Diabetes Risk Prediction Using Patient Health Data

1. Introduction

Diabetes mellitus is one of the common long-term illnesses in the world. It is also a reason for heart disease, kidney failure and other serious health problems. Finding out about diabetes mellitus early and predicting the risk of diabetes mellitus accurately are very important. They help give care lower long-term health costs and make life better for people who have diabetes mellitus. I believe that people are using intelligence, especially machine learning and deep learning increasingly to predict diseases with patient data. Classic machine learning models can predict diabetes mellitus well when they use everyday health measurements, like glucose, body mass index, blood pressure, age and other health signs [2,3].

Recently quantum machine learning has become a research direction for classifying data. Quantum machine learning is based on the idea that quantum feature spaces, which we reach by encoding data into circuits can show complex data patterns and make different groups easier to separate than normal feature spaces [8,9]. Among many quantum machine learning methods, the variational quantum classifier attracts a lot of attention. It mixes quantum circuits with classical optimization to do supervised classification [10,15].

2. Related Work

Predicting diabetes from clinical checks has been a hot topic in machine learning for over thirty years. The Pima Indians Diabetes dataset, first studied by Smith et al. With the ADAP adaptive learning algorithm [1] has grown into a go-to benchmark for testing classification models in this field. Many later papers have tried an array of classic algorithms—logistic regression, k-nearest neighbors, decision trees, support vector machines and ensemble tools such as random forests and gradient boosting—usually reporting accuracies between 0.70 and 0.97 depending on how data is cleaned and features are chosen [2,3,4]. Ensemble and mixed models often turn out to be classic baselines. Laila et al. Reached high as 97% accuracy with a Random Forest on UCI early-stage diabetes data [3] while Ahmed et al. Hit 94.87% using a blend of support-vector, artificial-neural-network and fuzzy-logic approaches [2]. Kurbanov et al. Offer a comparison of classic machine-learning methods, for diabetes prediction showing that tree-based ensembles stay hard to beat [4].

Beyond table tests, classical machine learning has grown to handle large clinical and administrative data sets. Ravaut et al. Predicted diabetes complications from the health records of more than one million patients reaching an AUC of 0.777 [6] and electronic-health-record (EHR) gradient-boosting frameworks such as LightGBM have achieved AUC values as high as 0.969 [7]. Systematic reviews of machine learning for diabetes management in care [5] and meta-analyses of artificial intelligence and machine learning for patient-data management in related chronic conditions such as hemodialysis [23] highlight both the promise and the reproducibility challenges of these approaches. Related work on predicting life expectancy from lifestyle factors [22] and on deep-learning-based patient monitoring [21] further illustrates the breadth of machine learning applications, across clinical and lifestyle health data.

In parallel quantum machine learning has become a way to solve classification tasks that could use richer feature representations. Havlíček et al. Showed that quantum circuits can transform inputs into high-dimensional Hilbert spaces and they proposed two supervised methods: a quantum kernel estimator that works with a classical support-vector machine and a variational quantum classifier that learns directly on the quantum feature map [8]. Killoran clarified the link between quantum feature maps and kernel methods proving that data encoded into a quantum state is automatically placed in a feature Hilbert space where inner products give a kernel [9]. These basic results directly inspire the Quantum Support Vector Classifier (QSVC) and Variational Quantum Classifier (VQC) that we evaluate in this study.

A growing number of studies have used these methods for diabetes prediction with the Pima Indians dataset. The evidence that appears however shows that quantum models usually match, than beat well-tuned classical baselines. Gupta et al. Compared quantum machine learning with learning and found 86% accuracy for the quantum model versus 95% for a classical deep network [10]. Vimal obtained 74% accuracy with a QSVC compared with 76% for a SVM on the same Pima data [13] and Genç found that a quantum SVM reached 84.3% versus 85% for logistic regression on the larger BRFSS Diabetes Health Indicators dataset [14]. Maheshwari et al. Reported that an ensemble Qboost classifier achieved 69% accuracy but with a 55-fold speed-up over the classical pipeline [11] while Prakash et al. Reported accuracies high as 97% for a quantum algorithm on a diabetes-mellitus dataset [12]. The best quantum results come from tuned variational circuits: Khan et al. Examined 76 VQC variants across four- and eight-qubit configurations and found that the best model—six layers with a ZZ feature map and a COBYLA optimizer—reached 85% accuracy on eight qubits (versus 80% on four) beating SVM, Random Forest, Decision Tree and Logistic Regression baselines [15]. Quantum-inspired models that run entirely on hardware have also shown strong performance; the QISICGM framework of Young reported an F1-score of 0.893 and an AUC of 0.870 on augmented Pima data surpassing Random Forest and XGBoost [16]. Table 1 summarizes representative comparisons, for diabetes prediction.

Two limitations keep appearing in this research. First quantum models for predicting diabetes have mostly been tested on the Pima dataset. They have not been checked on datasets in real hospital environments or using large electronic health records [10,15]. This is a problem that also affects classical models in the same area [5]. Second, the choice of feature maps the optimizer used, the depth of the quantum circuit and the number of qubits all have an impact on how well a quantum model performs [15]. This makes it hard to compare quantum models fairly. The current study helps fill this gap by providing a controlled comparison between QSVC and VQC and three classical models. It uses datasets, preprocessing steps and evaluation methods for all models. This adds to the number of studies that can reliably compare quantum and classical approaches, for predicting diabetes risk.

3. Materials and Methods

3.1. Dataset Overview

The dataset used in this study is called diabetes.csv which was created by the National Institute of Diabetes and Digestive and Kidney Diseases and shared by Smith et al. [1]. This dataset has 768 records each one showing a person with different medical and personal details that are connected to diabetes risk. Each record has eight features that can predict something and one target variable that’s. Yes or no as shown in Table 2. All the people in the dataset are female, at 21 years old and belong to the Pima Indian group.

3.2. Data Preprocessing

Several physiological features in the dataset—Glucose, BloodPressure, SkinThickness, Insulin, and BMI—contain zero entries that are not biologically plausible and therefore represent missing values. These entries were identified as missing and addressed during preprocessing. Because both the margin-based classical models and the quantum feature-encoding models are sensitive to the magnitude of the input features, all predictors were standardized to a comparable scale before training; feature scaling is especially important for the quantum classifiers, whose ZFeatureMap encodes each feature as a single-qubit rotation angle. The processed data were then partitioned into training and test subsets, and the identical split was used for every classical and quantum model to ensure a fair comparison.

3.3. Classical Machine Learning Models

Three classical baselines were implemented with the scikit-learn library [17] to establish performance benchmarks. Each classifier was configured with standard hyper-parameters chosen to limit overfitting on the relatively small dataset, as described below.

3.3.1. Model Architectures

From my perspective the logistic regression model acts as a classifier that converts input features into a probability of a class using the logistic function. The logistic regression model was regularized by setting an inverse-regularization strength of C = 0.1 solved with the routine and limited to one hundred iterations. The decision tree classifier splits the feature space by using the Gini impurity criterion. Was capped at a maximum depth of five with at least five samples required to split an internal node, which helps prevent overfitting. The random forest classifier is a group of one hundred decision trees; each decision tree grows to a depth of ten and considers the square root of the total number of features at each split. By averaging predictions, from decorrelated decision trees the random forest classifier improves generalization and as reported in Section 4, the random forest classifier provides the strongest classical baseline in our experiments.

3.4. Quantum Machine Learning Models

Quantum models were implemented using the Qiskit Machine Learning library [18,19], leveraging quantum kernels and parameterized circuits for binary classification:

3.4.1. Quantum Support Vector Classifier (QSVC)

The QSVC takes every record and turns it into a quantum state using a ZFeatureMap. This ZFeatureMap uses single-qubit rotations that are controlled by the input features. These rotations help move the data into a dimensional quantum Hilbert space. After that a quantum kernel is found by looking at how much the feature-mapped states overlap with each other. This kernel matrix is then given to a support-vector machine. The support-vector machine uses this information to find the decision boundary. The number of qubits used is the same, as the number of features that are encoded. The feature map uses a number of repetitions to manage how deep the circuit is. Since it is very hard to simulate the kernel the QSVC was trained on a sample of the training data.

3.4.2. Variational Quantum Classifier (VQC)

The Variational Quantum Classifier follows the paradigm that was introduced by Havlíček and colleagues in reference [8]. The VQC first takes inputs and embeds them using the same ZFeatureMap. Then the classifier applies a trainable RealAmplitudes ansatz, which is a hardware-efficient circuit that alternates single-qubit rotations and entangling gates. The Variational Quantum Classifier uses one repetition of the ansatz to keep the circuit shallow and to avoid trainability problems such as barren plateaus. After measurement the classifier maps the outputs to class labels. The VQC then uses the gradient-free COBYLA optimizer to adjust the circuit parameters and minimize the classification loss. As with the QSVC the VQC performed training on a reduced subset of the data because of simulation constraints.

3.5. Experimental Setup and Evaluation Metrics

All models were trained and evaluated on the same training and test partitions. The three classical baselines were implemented in scikit-learn [17] with the hyper-parameters described in Section 3.3, whereas the two quantum classifiers were implemented in Qiskit Machine Learning [18,19] and executed on a classical statevector simulator rather than on physical quantum hardware. Because simulating the quantum kernel and the variational circuits is computationally expensive, the QSVC and VQC were trained on a reduced subset of the training data, as noted in Section 3.4.1 and Section 3.4.2.

Model performance was assessed using four standard classification metrics—accuracy, precision, recall, and the F1-score—computed from the confusion-matrix counts of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN):

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

$$F_1-score={\displaystyle \frac{2\times Precision\times Recall}{Precision+Recall}}$$

Here the positive class corresponds to patients diagnosed with diabetes (Outcome = 1). Because the dataset is class-imbalanced, with negative cases outnumbering positive ones, the F1-score is reported alongside accuracy as a more informative measure of performance.

4. Results

This section reports the performance of the classical and quantum classifiers on the held-out test set, organized by model family and concluding with a direct comparison of all five models.

4.1. Classical Model Performance

Table 3. Accuracy of the classical machine learning models.

Model	Accuracy	Precision	Recall	F1-Score
Logistic Regression	0.72	0.58	0.70	0.64
Decision Tree	0.68	0.56	0.46	0.51
Random Forest	0.73	0.62	0.59	0.60

Table 3. Accuracy of the classical machine learning models.

Model	Accuracy	Precision	Recall	F1-Score
Logistic Regression	0.72	0.58	0.70	0.64
Decision Tree	0.68	0.56	0.46	0.51
Random Forest	0.73	0.62	0.59	0.60

4.2. Confusion Matrices

All classical machine learning models were evaluated using confusion matrices, as shown in Figure 1.

4.3. Quantum Model Performance

Table 4. Accuracy of the quantum machine learning models.

Model	Accuracy	Precision	Recall	F1-Score
QSVC	0.66	0.65	0.60	0.62
VQC	0.65	0.60	0.58	0.59

Table 4. Accuracy of the quantum machine learning models.

Model	Accuracy	Precision	Recall	F1-Score
QSVC	0.66	0.65	0.60	0.62
VQC	0.65	0.60	0.58	0.59

4.4. Comparative Performance

The performance metrics for all models were compiled into a single table for comparison (Table 5).

5. Discussion

The comparative analysis of the applied models shows key points. The classical Random Forest classifier delivered the accuracy at 0.73 and neither quantum model was able to outperform it. The QSVC achieved an accuracy of 0.66. This quantum approach seemed to capture some -linear relationships between features like glucose, BMI and age. However, the decision boundary created by QSVC was not strong enough to compete with the tree-based method. A major limitation is that QSVC was trained on a portion of the data due to simulation limits. So, its performance should be seen as a bound and not an exact measure. The VQC reached an accuracy of 0.65. This result is like classical models but still lower than the Random Forest. Its performance appears limited by factors: the circuit depth is shallow the number of qubits is small and issues like barren plateaus made optimization difficult. Across all models’ precision values were generally lower than recall values. This trend was especially clear in the Decision Tree and Logistic Regression classifiers. This pattern reflects the class imbalance in the dataset—there are more negative cases than positive ones. In total neither quantum model beat the classical baseline on this dataset. The QSVC performed better than the VQC. Both fell short compared to the Random Forest. These results match studies that show quantum models tend to match, rather than surpass well-optimized classical models when predicting diabetes, in the current NISQ era [10,13,14].

6. Conclusions

This study carried out a comparison between machine learning algorithms, such as Logistic Regression, Decision Trees and Random Forests and new quantum-enhanced classifiers namely the Quantum Support Vector Classifier (QSVC) and the Variational Quantum Classifier (VQC) all used for classifying diabetes cases. Our findings indicate that although classical ensemble methods, Random Forest still set strong performance standards quantum-hybrid models show promise when data dimensions are reduced. We saw that VQC and QSVC can learn non-linear decision lines inside high-dimensional Hilbert spaces highlighting their ability to act as feature-mapping tools. However, the gap in performance between quantum simulations run on classical machines and well-known classical deep learning systems points to serious difficulties. These problems mainly come from limits on data encoding and, from the size of quantum feature maps that exist during the Noisy Intermediate-Scale Quantum or NISQ era.

7. Future Work

To push forward the integration of quantum machine learning into diagnostic applications, future work will follow three paths. The first path is featuring encoding. We will look at expressive quantum embeddings, such as angle or amplitude encoding to raise the amount of information stored in quantum states and to close the gap with classical deep-learning models. The second path is hybrid optimization. We will combine powerful classical optimizers with hybrid training methods—for example quantum natural gradient descent - to cut the time needed for convergence and to make VQC training more stable on larger data sets. The third path is real-hardware benchmarking. We will move from simulation to running on real quantum processing units (QPUs). This will let us see how decoherence, gate noise and measurement error affect classification accuracy and give a picture of quantum advantage in noisy diagnostic situations. Finally, we must validate these models on groups and, on large electronic health records not just the Pima dataset, which is an important next step. Validating these models on external cohorts and large-scale electronic health records, beyond the Pima dataset, remains an important further step [15].