Machine Learning System Personalizes Clinical Test Recommendations for Individual Patients

A team of researchers has built a machine-learning system called iCARE that does something straightforward but potentially powerful: it figures out which medical test to order next for each individual patient, rather than recommending the same test to everyone.

The problem iCARE addresses is real. In clinical practice, doctors often follow standard diagnostic pathways—the same sequence of tests for most patients. But patients are not identical. A test that reveals crucial information for one person might be redundant or uninformative for another. The researchers reasoned that if a machine could learn from the medical histories of past patients, it could identify patterns in which tests matter most for which kinds of people, then apply those patterns to new patients arriving in the clinic.

The framework works in three steps. First, it calculates how similar a new patient is to all the patients in a historical database, using a mathematical measure called Euclidean distance. Patients who resemble the incoming person more closely receive higher weights. Second, it trains a weighted logistic regression model—a statistical tool that learns to predict disease or health status—using those weighted samples. This ensures the model focuses on patients most like the one being evaluated. Third, it uses a technique called SHAP (Shapley Additive Explanations) to determine which features, or data points, matter most for that specific patient's prediction. If an important feature is missing from the patient's current test results, iCARE recommends collecting it.

The researchers tested iCARE on both synthetic datasets designed to show how the system should behave under ideal and challenging conditions, and on real medical data. They used three datasets from the UCI Machine Learning Repository: one tracking early-stage diabetes risk with 16 features including age, gender, and symptoms like excessive thirst and frequent urination; one on heart failure with 13 clinical attributes including age, anemia status, and ejection fraction; and one on heart disease with 14 features including chest pain type and cholesterol levels. In each case, they simulated a realistic scenario where patients arrived with only a few initial test results and the system had to recommend which additional tests would be most informative.

On the early diabetes dataset, iCARE showed clear advantages. When patients came in with just three initial features—age, gender, and obesity status—the global approach (which recommends the same test to everyone) suggested polydipsia, or excessive thirst, 75 percent of the time. But iCARE recommended polyuria, or frequent urination, for 68 percent of patients and polydipsia for 32 percent. Both are classic diabetes symptoms, but the framework recognized that different patients benefit from different tests depending on their profile. The personalized approach improved accuracy and area under the curve (AUC), a measure of diagnostic performance, by 6 to 12 percent compared to standard methods. These improvements were statistically significant.

However, the heart failure dataset told a different story. iCARE showed no meaningful advantage over the global approach. The researchers attribute this to the structure of the data itself: in that dataset, the additional features available for recommendation did not have distinct predictive power for different patient subgroups. When all patients benefit equally from the same next test, personalization adds no value. This finding matters because it reveals an important limitation: iCARE works best when heterogeneity exists—when different patient populations genuinely benefit from different tests.

The researchers also compared iCARE to other feature selection methods, including sequential forward selection, LASSO regularization, and an imputation-based approach called eGuided. On the early diabetes dataset, iCARE achieved a 6 percent higher AUC score on average than eGuided, and on the heart disease dataset, 12.1 percent higher. The timing analysis showed that iCARE runs faster than most alternatives when the number of features is large, though the global SHAP method remains quickest overall.

The work carries real limitations. The framework requires a pool of historical patient records with complete feature sets to work effectively. It assumes that the initial features a patient has are informative about which additional features will be predictive—an assumption that may not always hold. The researchers tested on publicly available datasets rather than actual hospital data, which means real-world performance remains unknown. They also note that iCARE currently lacks an automated way to determine whether a given dataset will actually benefit from personalization, requiring researchers to run experiments to find out. Before deployment in clinical settings, the system would need validation on real hospital data and compliance with privacy regulations like HIPAA and GDPR. The authors acknowledge security risks inherent in training models on sensitive patient information and flag the need for fairness-aware methods to ensure recommendations work equitably across diverse populations.