Advancing clinical prognosis and diagnosis with longitudinal health data and time-dynamic prediction models (360G-Wellcome-320151_Z_24_Z)

Clinical prognosis and diagnosis of diseases rely on clinical prediction models that predict health outcomes with patient clinical characteristics. Improving clinical prediction model performance has the potential to direct more accurate delivery of treatment and intervention before disease onset, and identifying disease-susceptible individuals or disease-inducing clinical characteristics. Despite the increasing availability of clinical longitudinal data (data measured repeatedly over time for each patient), there is no uniform or credible scheme for adapting clinical prediction models based on longitudinal health data. The temporal structure of longitudinal health data provides both opportunities and challenges to the development of clinical prediction models. In this project we aim to improve clinical prognosis and diagnosis by leveraging recent advances in probabilistic machine learning to empower clinical prediction models with clinical longitudinal data. We anticipate a model architecture which, instead of giving a static prediction of patient clinical outcome, re-evaluates and updates its prediction each time the patient provides a new clinical record by evaluating the patient’s observable clinical histories. We envision this model to assist in providing more personalised, accurate and well-calibrated clinical risk prediction to the patient community and provide new insights on the long-term development of diseases reflected upon clinical history.