RESEARCH OVERVIEW
I am researching ways to ensure that machine learning models are robust to data distribution changes after deployment in healthcare.
Patient information recorded in electronic health records (EHR) are recorded as asynchronous events. For machine learning models, it is often convenient to convert these representations into synchronous blocks of time. This transformation results in discretised data with exceptionally high amounts of missing data (i.e. blood samples are not drawn hourly). I am interested in ways to represent patient trajectories as a single vector in a robust manner so that we can fine tune models for automated clinical decisions, ensure performance when transporting models across multiple institutes, and create decision performance that endures the chaotic evolution of medicine.
Relevant literature:
- Che, Zhengping, et al. "Recurrent neural networks for multivariate time series with missing values." Scientific reports 8.1 (2018): 6085.
- Song, Huan, et al. "Attend and diagnose: Clinical time series analysis using attention models." Thirty-Second AAAI Conference on Artificial Intelligence. 2018.
- Zemel, Rich, et al. "Learning fair representations." International Conference on Machine Learning. 2013.
- Alsentzer, Emily, et al. "Publicly available clinical BERT embeddings." arXiv preprint arXiv:1904.03323 (2019).
- Purushotham, Sanjay, et al. "Variational recurrent adversarial deep domain adaptation." (2016).
For a list of my publications, please check my Google Scholar page.