Grant: Travel Awards - Institute Community Support

Machine Learning (ML) is a current buzzword in many public, commercial and scientific domains. This approach has given rise to advancements in speech recognition, medical imaging, precision medicine, and drug development due to the rapid developments in computing power and increasing availability of data. Current ML applications in clinical epidemiology include identification of risk factors for health outcomes, classification of patients into risk strata to optimize individual treatment recommendations, and surveillance over the course of treatment. This is a promising avenue in modern HIV research as >36.9M people are living with HIV worldwide. The popularity of ML coincides with the implementation of electronic medical records (EMR) in healthcare systems globally. However, EMRs containing HIV patient data vary highly in content, scope, and quality across different health jurisdictions and institutions. Yet they are being used to develop ML algorithms to inform clinical decision making or to gain evidence regarding the effectiveness and safety of HIV treatments at the population level. The enthusiasm of automated learning from big data brings up risks of disregarding fundamental principals of proper study design and data collection, rendering findings invalid due to confounding bias: association does not imply causation! Over the last 30 years, there has been tremendous success in identifying the conditions necessary for causal claims based on observational data. The ML research community is starting to recognize the fundamental importance of these developments, which is igniting a major paradigm shift from associational to causal ML approaches. Applying the causal inference framework, we evaluate current structures of HIV EMR data in the literature and from a large HIV cohort based in Montreal to assess threats to the validity of inferences and outcome predictions (e.g. treatment adherence) using machine learning methods (random forests, neural networks).