55 APPLYING MACHINE LEARNING TO DIFFERENTIATE PREDICTORS OF SYNCOPE, SIMPLE AND COMPLEX FALLS IN THE IRISH LONGITUDINAL STUDY ON AGEING (TILDA)

Abstract

Background: Falls in older adults are clinically heterogeneous and can be classified as simple (accidental), complex (recurrent, unexplained, injurious), and due to syncope. The aim of this study was to generate machine learning models for these fall types in TILDA using the Syncope-Falls Index (SYFI), a 40-deficit index covering a wide range of risk factors (Fitzpatrick & Romero-Ortuno, 2021). Methods: New self-reported events of simple, complex falls, and syncope were recorded in participants two-yearly between Wave 1 (2010) and 4 (2016). The 40 SYFI features, age and sex, were entered into three separate random forest models. The dataset for each model was balanced by equalising the number of participants who did and did not have each event. Feature importances were derived and those with scores of ≥0.05 are reported below. Results: For simple falls (217 events, balanced dataset N=434), the most important predictors were age (0.13), pre-existing hypertension (0.07), presence of urine incontinence (0.07), polypharmacy (0.06) and MMSE<24 (0.05) (model accuracy: 0.53). For complex falls (1,077 events, balanced dataset N=2,154), top predictors were age (0.14), gender (0.07), osteoporosis (0.07), osteoarthritis (0.06) and unsteadiness getting up from a chair (0.05) (accuracy: 0.58). For syncope (185 events, balanced datasetN=370), top predictors were age (0.16), osteoporosis (0.07), unsteadiness getting up from a chair (0.06) and previous myocardial infarction (0.05) (accuracy: 0.50). Conclusion: In keeping with the literature, advancing age was the most important feature in all models. Further, a more nuanced understanding emerged highlighting possible roles of reduced cognition in simple falls, musculoskeletal disease in complex falls, and orthostatic intolerance and heart disease in syncope. Many predictors identified as relevant are modifiable, and this ranked approach could help prioritise interventions for maximum population benefit. The accuracy of the models however was moderate, underscoring the known difficulty in predicting falls even when attempting to break their clinical heterogeneity.