Recent work has shown that fine-tuning large networks is surprisingly sensitive to changes in random seed(s). We explore the implications of this phenomenon for model fairness across demographic groups in clinical prediction tasks over electronic health records (EHR) in MIMIC-III —— the standard dataset in clinical NLP research. Apparent subgroup performance varies substantially for seeds that yield similar overall performance, although there is no evidence of a trade-off between overall and subgroup performance. However, we also find that the small sample sizes inherent to looking at intersections of minority groups and somewhat rare conditions limit our ability to accurately estimate disparities. Further, we find that jointly optimizing for high overall performance and low disparities does not yield statistically significant improvements. Our results suggest that fairness work using MIMIC-III should carefully account for variations in apparent differences that may arise from stochasticity and small sample sizes.

The ability to track mental health conditions via social media opened the doors for large-scale, automated, mental health surveillance. However, inferring accurate population-level trends requires representative samples of the underlying population, which can be challenging given the biases inherent in social media data. While previous work has adjusted samples based on demographic estimates, the populations were selected based on specific outcomes, e.g. specific mental health conditions. We depart from these methods, by conducting analyses over demographically representative digital cohorts of social media users. To validated this approach, we constructed a cohort of US based Twitter users to measure the prevalence of depression and PTSD, and investigate how these illnesses manifest across demographic subpopulations. The analysis demonstrates that cohort-based studies can help control for sampling biases, contextualize outcomes, and provide deeper insights into the data.