top of page
Current Research
I am broadly working at the interface of classical mathematical modelling and machine learning. I am a firm believer that the combination of principle driven and data driven approaches is a powerful approach to almost any real world problem. One such problem is finding optimal treatment strategies for critical care patients, which I am focusing as my PhD research topic. I'm working towards translating physiological knowledge using mechanistic approaches, and then hoping to use it to extract high quality, interpretable treatment strategies using (Batch) Reinforcement Learning for critical care areas, with no agreed upon treatment strategies. One such example is vasopressor and IV fluid treatment for sepsis, where medical research has shown significant variation on treatment responses among patients, and there's little agreement on the best practices, despite being attributed to as high as 35% of all hospital deaths in USA in the past.
Precision medicine, individualized treatment tailored to an individual has been proposed as a promising approach to treat icu patients, especially in the case of Sepsis, where it has been observed that responses to treatment has been varies dependent on individual. Modern machine learning approaches and increased volume of patient data such as the MIMIC project, has opened up the possibility of better understanding individual patient needs, and treatment responses. However there are many challenges in applying standard Deep Reinforcement Learning techniques directly to the healthcare domain, for instance online data collection is prohibitive, and the data is essentially limited. Also safety is obviously critical, so using mechanistic knowledge to infer physiological parameters could help in deriving safe and interpretable treatment strategies.
Unifying Cardiovascular Modelling with Deep Reinforcement Learning for Uncertainty Aware Control of Sepsis Treatment
In this work we use a novel physiology driven recurrent autoencoder to infer patient specific cardiovascular states, and combine Deep RL, with Uncertainty Quantification, to propose a safe, uncertainty aware medical support system for septic patients.
Now published in PLOS Digital Health : https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000012
Preprint: https://arxiv.org/abs/2101.08477
Github Repo: https://github.com/thxsxth/POMDP_RLSepsis
Challenges in Applying Deep Reinforcement Learning for Critical Care Applications
ICCAI 2021 AI in Critical Illness: Emergence and Emergent Issues : Selected as one of the 10 best abstracts for presentation
Will be published in Journal of Critical Care as an extended abstract.
Inferring Patient Specific Cardiovascular States by Physiology Driven Self Supervised Learning
ICCAI 2021 AI in Critical Illness: Emergence and Emergent Issues : Selected as one of the 10 best abstracts for presentation
Will be published in Journal of Critical Care as extended abstract.
Other Research Interests
-
Offline Reinforcement Learning
-
Multi-Task RL and Meta Reinforcement Learning
-
Inverse Problems in Dynamical Systems and ODEs
-
Bayesian Inference
-
Mathematical Finance
Selected Previous Projects
-
A Survival based Machine Learning Approach to identify key bio-markers governing Sepsis Mortality.
-
Santander Customer Transaction Prediction
-
Kaggle competition (Top 5% Finish)
-
-
Speech to Text using Listen Attend Spell and Attention
-
Face classification and Verification using CNNs
To see more or discuss possible work let's talk >>
bottom of page