Fall risk prediction and validation in older adults Leveraging electronic health records with machine learning

Falls in older adults are common and comorbid. Many fall risk stratification tools exist but they are often time-consuming and have limited accuracy. The goal of this thesis is to leverage machine learning and EHR data to create tools that can reliably estimate an individualized fall risk and provide a sound basis for decision-making regarding interventions to reduce fall risk.
We performed a systematic review to compare falls prediction models in community-dwelling older adults based on routinely-collected data (RCD) and data from research cohorts. Both types of models performed similarly, with RCD-based models having an advantage in data collection efficiency. However, many models had limitations and biases that affect their reliability. Additionally, we developed and validated falls prediction models using EHR data in primary care and hospital settings. These models demonstrated fair discrimination and reasonable calibration. Furthermore, we applied Natural Language Processing (NLP) to general practitioners' (GPs) unstructured clinical notes in two distinct applications. The first application revealed that these notes can enhance falls prediction models when combined with structured clinical data. The second application identified topics associated with falls and their trends over time. Many topics, such as medications, showed increasing trends, serving as warning signs for imminent falls. Recognizing these trends early is crucial for proactive interventions to reduce fall risk.
Finally, we provided an overview of the SNOWDROP project, which aimed to create a data-driven approach for personalized fall risk prediction to support joint medication management between older adults and GPs. The resulting prediction models can facilitate shared decision-making, potentially reducing the risk of falls in clinical practice.