Abstract Details

Previous research has utilized machine learning models to quantify complex, non-linear relationships between clinical and demographic characteristics and stroke-related outcomes. However, very little research has been dedicated to applying these predictive techniques to quantify exact time-to-death relationships, especially in the context of specific interventions. These techniques can be leveraged to identify how treatment and control groups differ in their predictive relationships between clinical variables and mortality risk.

This study aimed to establish multivariate predictive models for quantifying survival times after stroke/transient ischemic attack (TIA) in the context of a clinical intervention study (PREMIERS) targeting risk reduction through periodontal disease treatment. A secondary aim was to explore differences in predictors between treatment and control groups, given that the treatment group showed beneficial, albeit nonsignificant, improvements in major adverse cardiovascular event recurrence risk.

Patients with stroke/TIA and moderate periodontal disease were randomly assigned to either standard or intensive periodontal treatment arms (140 standard, 140 intensive). At baseline, study participants were assessed for medical history, stroke severity, vital signs, demographics, blood tests, and carotid artery stenosis. Random Survival Forest (RSF) analysis was applied to examine multivariate predictive relationships between baseline variables and time-to-death in the overall study group and intensive and standard treatment arms separately.

RSF models showed that survival time could be effectively predicted (C>0.5) for the overall study group (C=0.35) and the intensive treatment arm (C=0.42), but not the standard treatment arm (C=0.66). Blood pressure was found to be a significant predictor overall, while stroke severity and recurrence risk was uniquely predictive in the intensive treatment group.

Machine learning may be useful in intervention to pinpoint baseline variables which uniquely predict mortality risk, depending on treatment assignment. Future applications of such models could also generate individual estimates of survival probability trajectories to provide personalized treatment recommendations.