A machine learning and pathfinding framework reduced cumulative cardiovascular disease risk modification costs from 0.99 to 0.88 versus random paths (P < 0.01).
Does a machine learning and pathfinding framework reduce the cumulative cost of lifestyle modification pathways compared to randomly generated pathways in patients at risk for cardiovascular disease?
A novel machine learning and pathfinding framework successfully identified optimal lifestyle modification pathways with lower cumulative costs compared to random pathways, offering a potential decision support tool for cardiovascular disease prevention.
Absolute Event Rate: 0% vs 0%
Abstract Background Data-driven personalized treatment approaches enhance shared decision-making for cardiovascular disease prevention. Conventional models identify key health factors but struggle to optimize behavioral change sequences and maintain patient motivation. We hypothesized that integration of machine learning and pathfinding algorithms could simultaneously derive optimal intervention pathways and cost structures to sustain patient adherence. Purpose To develop a framework integrating machine learning and pathfinding algorithms that creates optimal lifestyle modification pathways while minimizing costs, improving long-term adherence to cardiovascular risk reduction goals. Methods Our three-step framework included: (1) XGBoost model development to assess cardiovascular disease probability and identify its two most important modifiable risk factors; (2) creation of a grid-world environment with these two key modifiable risk factors as axes, where edge costs between patient states were weighted by cardiovascular disease probability changes (improvements received lower costs); and (3) application of breadth-first search to identify optimal pathways minimizing costs to the node with the most improved cardiovascular disease probability (Fig.1). The model was validated using a publicity-available combined dataset comprising five well-known heart disease cohorts (Statlog Heart, Cleveland, Hungarian, Switzerland, and Long Beach), featuring 11 clinical factors and 1 target response indicating heart disease presence. Model performance of the proposed framework was compared with randomly generated pathways (Control) using a t-test. Results From 1,018 eligible individuals in the combined dataset, an XGBoost model trained with 814 participants (80%) and tested with 204 (20%) achieved an F1-score of 0.92, identifying systolic blood pressure and total cholesterol as the two key modifiable risk factors. Among the 204 test samples, 85 had a disease probability ≥50% with a true positive label; 83 patients aged ≥40 were used for the optimal pathfinding experimentation. Average cumulative cost in the proposed pathfinding framework was lower than that in the control method using randomly generated pathway (0.88 vs. 0.99; P 0.01) (Fig.2). Conclusions We developed a novel pathfinding framework that had a potential to identify optimal paths to minimize costs for reaching desired disease risk levels. By incorporating cost structures into the pathfinding process, our framework supports continued behavioral change. This approach could improve decision support tools for treatment planning in cardiovascular disease prevention strategies.Fig 1 Fig 2
Kosaka et al. (Sat,) reported a other. A machine learning and pathfinding framework reduced cumulative cardiovascular disease risk modification costs from 0.99 to 0.88 versus random paths (P < 0.01).