We invite submissions from all areas of machine learning for health and biomedicine but we especially encourage submissions that focus on this year’s theme: What makes machine learning in medicine different?
Example focus areas include, but are not limited to, the following:
Data and Labels are Noisy and/or Missing
Healthcare data suffers from very high rates of noise and missingness, both in terms of input data as well as in terms of labels. Therefore, methods such as noise-robust models, data imputation, semi-supervised learning, distant/weak supervision, and clinician in the loop learning, all have an important role to play in machine learning for healthcare.
Causality / Confounding
Causality is central to any interventional task within healthcare, and methods to learn causal structures from observational data or to account for additional confounding within data are critical in healthcare.
Do No Harm: Trust, Generalizability, Interpretability, & Reproducibility
Machine learning models in healthcare need to demonstrate robustness, pass a high bar of trust before deployment and be routinely validated. This places strong burdens on tests of generalizability (to assert the model will extend to novel patients), interpretability (to provide auxiliary arguments in favor of the models’ correctness), and reproducibility (to ensure the model can be trusted, audited, and debugged).
Deployment Challenges within Healthcare
Deploying a machine learning model in healthcare faces significant hurdles and requires additional socio-technical validation and control mechanisms. In addition, models models deployed in healthcare may face regulatory challenges.
Healthcare policy, new clinical knowledge or even the deployment of models to production can fundamentally change the process machine learning algorithms intend to model. For example, numerous examples have shown that when CMS increases the reimbursement for a particular diagnosis, the diagnosis is used in claims more frequently.
Fairness & Bias
Fairness concepts are important in healthcare, for example in ensuring all patients receive unbiased treatment, and especially difficult to respect, as protected subgroup information is often a valid causal link to clinical state.
Multi-modality, High-dimensionality, "p >> n"
Healthcare data is often highly multimodal, extremely high-dimensional, and can have many more features than samples (e.g., genomics). Each of these aspects requires specialized methodological innovation.
Healthcare data is extremely sensitive and is therefore protected by specialized laws such as HIPAA. A key goal of medical ML is therefore the development of algorithms that can verifiably preserve patient privacy with minimal tradeoffs in performance.