IYSC14 - Using a Supervised Learning Algorithm to Predict Post-Operative Treatment Failure in Patients with Cardiovascular Disease via Passive Monitoring

Objectives: Peripheral arterial disease (PAD) affects an elderly patient population characterized by multiple comorbidities, healthcare access barriers, and limited mobility. Each surgical intervention and hospitalization related to PAD is a stressor on the patient’s body, and some patients never fully recover, with continued loss in functional status. Other patients, however, are able to regain their prior functional status and even surpass it, maintaining pain-free mobility. A key question for physicians is what differentiates these patients besides their outcomes?

Methods: We explore a way to predict postoperative treatment failure with pre-operative factors as well as by passively monitoring patients’ recovery and walking patterns. The dataset used in this study includes 101 PAD and cardiac patients followed for 1 year remotely before and after a major cardiovascular intervention. True to this patient population, it is a heterogenous group with an average age of 68.9 years and is characterized by various comorbidities (37.4% diabetic, 87.9% with hypertension), medications, and levels of smartphone literacy (26.3% smartphone naive). 71% received vascular interventions for PAD and 29% received cardiac interventions. 116 total interventions (95 initial, 11 reinterventions) were performed and included endovascular and open surgeries.

The algorithm was given 20 pre-operative factors for each patient and included demographic data, common cardiovascular and diabetes medication, as well as cardiovascular health markers measured before and after surgery (Ankle Brachial Index and Ejection Fraction). Procedure details included the type of vascular or cardiac procedure performed, and whether the procedure was successful. Lastly, the dataset includes preoperative and postoperative patient-reported mobility assessments and passively acquired quantitative mobility measures. Relevant measures include max steps without stopping (MSWS), activity index (Ai), total steps and distance covered within six minutes (6 minute walk test), and other accelerometer data collected remotely from an activity app on the patient’s smartphone and smartwatch.

Results: Our supervised learning algorithm constructs a 4-level decision tree with a subset of the 20 previously described factors to predict whether a patient will require follow-up surgery within 1 year of their initial intervention. When run 1000 times, the top 5 factors chosen were pre-op ABI, caregiver dependence, procedure type, pre-op EF, and hyperlipidemia diagnosis. The model was analyzed with 10-fold cross validation and correctly predicted the need for reintervention with a mean accuracy of 0.90 (standard deviation 0.06).

Conclusions: This algorithm models how passively monitoring patients can supply data for accurate patient outcome predictions in even the most difficult to manage diseases.