Predicting morbidity by local similarities in multi-scale patient trajectories

Area of Science:

• Computational health informatics
• Predictive modeling in healthcare
• Biomedical data analysis

Background:

• Patient trajectories (PTs) capture temporal patient evolution but are underutilized in healthcare prediction.
• Existing predictive models often use static patient data, missing crucial temporal dynamics.
• Previous PT-based prediction studies relied on single-source data, limiting their scope.

Purpose of the Study:

• To develop a methodology for identifying local similarities in multi-source patient trajectories (PTs) to predict future morbidities.
• To validate the proposed methodology for predicting cardiovascular disease (CVD) risk in patients with diabetes.

Main Methods:

• Proposed a novel formal definition for PTs using sequences of longitudinal multi-scale data.
• Developed a dynamic programming methodology for identifying local alignments within PTs to predict future morbidities.
• Validated the approach on a dataset for predicting CVD in diabetic patients.

Main Results:

• Achieved a precision of 0.33, recall of 0.72, and specificity of 0.38 in predicting CVD occurrence in diabetic patients.
• Demonstrated the potential of using local similarities in longitudinal multi-scale PTs for morbidity prediction.

Conclusions:

• The proposed methodology for defining and aligning PTs is generic and applicable across clinical domains.
• The approach shows significant utility for secondary screening, particularly in predicting CVD risk for diabetic individuals.