Individualized machine-learning-based clinical assessment recommendation system

Abstract

Traditional clinical assessments often lack individualization, relying on standardized procedures that may not accommodate the diverse needs of patients, especially in early stages where personalized diagnosis could offer significant benefits. We aim to provide a machine-learning framework that addresses the individualized feature addition problem and enhances diagnostic accuracy for clinical assessments.Individualized Clinical Assessment Recommendation System (iCARE) employs locally weighted logistic regression and Shapley Additive Explanations (SHAP) value analysis to tailor feature selection to individual patient characteristics. Evaluations were conducted on synthetic and real-world datasets, including early-stage diabetes risk prediction and heart failure clinical records from the UCI Machine Learning Repository. We compared the performance of iCARE with a Global approach using statistical analysis on accuracy and area under the ROC curve (AUC) to select the best additional features. The iCARE framework enhances predictive accuracy and AUC metrics when additional features exhibit distinct predictive capabilities, as evidenced by synthetic datasets 1-3 and the early diabetes dataset. Specifically, in synthetic dataset 1, iCARE achieved an accuracy of 0.999 and an AUC of 1.000, outperforming the Global approach with an accuracy of 0.689 and an AUC of 0.639. In the early diabetes and heart disease dataset, iCARE shows improvements of 6-12% in accuracy and AUC across different numbers of initial features over other feature selection methods. Conversely, in synthetic datasets 4-5 and the heart failure dataset, where features lack discernible predictive distinctions, iCARE shows no significant advantage over global approaches on accuracy and AUC metrics. iCARE provides personalized feature recommendations that enhance diagnostic accuracy in scenarios where individualized approaches are critical, improving the precision and effectiveness of medical diagnoses.

Fig 1. Architecture of the iCARE framework.

Data were obtained from an incoming patient (I), and weights were generated for the pool of known cases in the Similarity Calculation Module (II). Using these sample weights, we generate a weighted logistic regression model for an incoming patient (III). SHAP values are then generated using a SHAP explainer for all the subjects in the pool of known cases (IV). The Feature Recommendation Module will then gather all the individual SHAP values and produce a recommendation if there is a missing feature that can be recommended to the patient (V).

Fig 2. Experimental workflow to evaluate the iCARE framework.

The figure above highlights the main experimental workflow to evaluate the iCARE framework against traditional global feature selection. This workflow produces two distinct approaches to generating recommendations, as shown by the Global (i.e., global feature selection) and iCARE (i.e., individualized feature selection) split in part I. In addition, there are two distinct approaches to training the inference model, as shown in part II, where the logistic regression model can be trained with or without sample weights (i.e., LW or no LW). This produces four approaches: Global, Global+LW, iCARE, and iCARE+LW.

Fig 3. Synthetic dataset 1.

Two 2D scatter plots displaying the relationship between the initial feature (x-axis) and the added feature (y-axis). The red dots represent negative samples (e.g., sick patients), while the blue dots represent positive samples (e.g., healthy patients). The left plot depicts added Feature 1, exhibiting predictive power for Initial Feature < 0.5, while random noise is observed in the shaded area above Initial Feature > 0.5. The right graph illustrates added Feature 2, demonstrating predictive power for Initial Feature > 0.5, with random noise observed in the shaded area below Initial Feature < 0.5.

Fig 4. Synthetic dataset 2.

Two 2D scatter plots, similar to Fig 3, showcase the relationship between the initial feature (x-axis) and the added feature (y-axis). The red dots represent negative samples (e.g., sick patients), while the blue dots represent positive samples (e.g., healthy patients). Notably, the predictive area in this dataset exhibits a non-linear pattern, suggesting a more complex relationship between the features.

Fig 5. Synthetic dataset 3.

2D scatter plots resembling Fig 3, depicting the relationship between the initial feature (x-axis) and the added feature (y-axis). The red dots represent negative samples (e.g., sick patients), while the blue dots represent positive samples (e.g., healthy patients). Notably, the left graph demonstrates predictive power for X < 0.7, while the right graph showcases predictive power for X > 0.3. The green-shaded region highlights an overlapping area (0.3 < X < 0.7) where both features possess equal predictive power.

Fig 6. Synthetic dataset 4.

Scatter plots depicting the relationship between the initial feature and the added feature, resembling the format of Fig 3. Notably, both the left and right graphs illustrate identical predictive regions.

Fig 7. Synthetic dataset 5.

Each scatter plot represents a different feature’s predictive power. The first scatter plot demonstrates strong predictive capability, while the other two plots depict features with limited predictive utility. This visualization underscores the scenarios where one feature overpowers the other features.

Fig 8. Performance summary of Synthetic Dataset 1 - 3.

Comparison of accuracy (left) and area under the curve (AUC) (right) across three synthetic datasets. Each bar group represents a dataset, with values indicated for both global and local weighted metrics. For Dataset 1, the accuracy stands at 0.689, 0.667, 0.999, 0.999 with an AUC of 0.639, 0.814, 0.999, 1.0. In Dataset 2, the accuracy stands at 0.551, 0.767, 0.632, 0.891 with an AUC of 0.584, 0.850, 0.687, 0.953. Dataset 3 accuracy stands at 0.914, 0.894, 0.998, 0.998, along with an AUC of 0.888, 0.974, 0.996, 0.998. This comparison highlights variations in performance across the different synthetic datasets that represent ideal scenarios.

Fig 9. Performance summary of synthetic dataset 4 - 5.

Comparison of accuracy (left) and area under the curve (AUC) (right) across Synthetic Datasets 4 and 5. Each bar group represents a dataset with performance metrics for both global and iCARE. For dataset 4, accuracy values obtained were 0.747, 0.810, 0.738, 0.792, and AUC values obtained were 0.781, 0.805, 0.764, 0.787. For dataset 5, accuracy values obtained were 0.811, 0.740, 0.774, 0.742, and AUC values obtained were 0.790, 0.815, 0.799, 0.811. These results reveal two distinct scenarios where iCARE learning fails to substantially improve global learning regarding feature addition and inference.

Fig 10. Early Diabetes dataset performance summary.

This figure illustrates the mean performance of the early diabetes dataset on different feature spaces on accuracy and AUC metrics, with global and local perspectives represented by blue/orange and green/red lines, respectively. Error bars at each data point represent the standard deviation from the mean. The line graphs the maximum number of features towards the ceiling, represented by the purple line. The ceiling model represented an ML model trained on all features.

Fig 11. Heart failure dataset performance summary.

This figure presents a comprehensive overview of mean accuracy and AUC metrics across various feature spaces on the heart failure dataset, offering insights into global and local perspectives depicted by blue/orange and green/red lines, respectively. Error bars show the standard deviation, while convergence towards the maximum features underscores notable trends.

Fig 12. Performance Comparison of Feature Selection Methods Across Three Datasets.

This figure presents a comparative evaluation of AUC-ROC and Accuracy across different feature selection approaches on three real-world datasets: (A) Early Diabetes, (B) Heart Failure, and (C) Heart Disease. The graphs illustrate the performance of five feature selection methods which are Global (blue), SLS (orange), LASSO (green), iCARE (red), and eGuided (purple), across varying feature subsets. The x-axis represents the number of selected features, while the y-axis shows the corresponding AUC-ROC and Accuracy scores, providing insights into the effectiveness of each method in optimizing predictive performance.

Fig 13. Timing performance of feature recommendation methods across varying sample sizes and feature dimensions.

Runtime comparisons for five feature recommendation methods: Global SHAP, SFS, LASSO, iCARE, and eGuided, under two experimental conditions. The left panel shows how the runtime scales with the number of patients (N = 20 to 100) using the Early Diabetes Dataset, where each patient begins with three known features. The right panel displays runtime behavior as the number of available features increases (1 to 60) using a synthetic dataset of 500 samples and 100 total features.