Research Projects
Evaluating XAI Explanations for Clinical Decision Support Systems
About Project:
Modern Artificial Intelligence (AI) models offer high predictive accuracy but often lack interpretability with respect to reasons for predictions, which is important in many high-stakes decisions. Many Explainable AI (XAI) techniques have been designed to generate explanations for predictions from blackbox models. However, there are no rigorous metrics to evaluate these explanations, especially with respect to their usefulness to clinicians. We develop a principled method to evaluate explanations by drawing on theories from social science and accounting for specific requirements of the clinical context.
Project supervised by Prof Vaibhav Rajan.
Publications:
Towards a Theory-Based Evaluation of Explainable Predictions in Healthcare, Suparna Ghanvatkar, Vaibhav Rajan, Short Paper, International Conference on Information Systems ICIS 2022, Copenhagen, Denmark. PDF
Evaluating Explanations from AI Algorithms for Clinical Decision-Making: A Social Science-based Approach, Suparna Ghanvatkar and Vaibhav Rajan, accepted at IEEE Journal of Biomedical and Health Informatics (J-BHI), DOI: 10.1109/JBHI.2024.3393719
Risk Prediction in ICU using heterogeneous data sources
About:
Risk prediction models in Intensive Care Units (ICU) are important for clinical decision-support tasks such as identifying high-risk patients and prioritizing their care. Different sources of patient information in ICUs can be utilized for such models: electronic health records record demographics, laboratory results, vitals, and clinical notes, and bedside monitors record very high-frequency data such as ECG. Most of the data is temporal but at different time scales. For example, vitals may be measured once in a few hours while ECG consists of 125 measurements per second. The available data is also multimodal, such as text nursing notes, radiology images, time-series data. We attempt to integrate data available at such varying time-scales and in different modalities to perform efficient risk prediction.
Project supervised by Prof Vaibhav Rajan.
Publications:
Deep Recurrent Neural Networks for Mortality Prediction in Intensive Care using Clinical Time Series at Multiple Resolutions, Suparna Ghanvatkar, Vaibhav Rajan, Short Paper, International Conference on Information Systems ICIS 2019, Munich, Germany. PDF
In-progress work: Graph-Based Patient Representation for Multimodal Clinical Data: Addressing Data Heterogeneity, Suparna Ghanvatkar, Vaibhav Rajan, WIP available at medrxiv: link
Past Research Projects
Personalized Profiles for Physical Activity Intervention
About:
The first step towards the personalization of physical activity interventions is the development of user-profiles. Our preliminary results indicate that temporal profile, i.e. one which considers sequence of activities of user, differentiates between two users with sedentary and active lifestyles more effectively than a non-temporal profile. As a next step, we evaluated the effect of timing of the intervention based on the preference of time of day according to circadian
rhythm. Our preliminary results indicate improvement as compared to interventions at fixed timings.
Project supervised by Prof. Atreyi Kankanhalli and Prof. Vaibhav Rajan.
Publications:
User Models for Personalized Physical Activity Interventions: A Scoping Review , Suparna Ghanvatkar, Atreyi Kankanhalli, Vaibhav Rajan, JMIR mHealth and uHealth, JMIR Mhealth Uhealth 2019;7(1):e11098. DOI: 10.2196/11098
Detecting Temporal Pattern Profiles from Smartphones for User Activity Analysis, Suparna Ghanvatkar, Vaibhav Rajan, Atreyi Kankanahalli, Short Paper, International Conference on Information Systems ICIS 2018, San Francisco, USA. PDF
Temporal Personalization of a Digital Intervention for Physical Activity, Suparna Ghanvatkar, Saurabh Chaudhari, Atreyi Kankanhalli, Short Paper, Pacific Asia Conference on Information Systems PACIS 2022. PDF
Personalization of Intervention Timing for Physical Activity: Scoping Review, Saurabh Chaudhari, Suparna Ghanvatkar, Atreyi Kankanhalli, JMIR Mhealth Uhealth 2022;10(2):e31327, DOI: 10.2196/31327
Necrosis Identification in Glioblastoma H&E Slides
Glioblastoma Multiforme is the most severe of brain tumours with a very low patient survival rate and very fast infiltration. Glioblastomas are not surgically curable, but patients have better prognosis with removal of tumour regions. The automated identification of key regions such as necrosis and endothelial proliferation can speed up the process for neuro-pathologists. We focussed on the task of identifying necrosis regions using a few manually annotated obtained during semester-long research internship at Mazumdar Shaw Medical Foundation, Bangalore, India.
Project supervised by Prof. Neelam Sinha.
Github: link
Automated Image Annotation from Weak Labels
The internet is full of images consisting of strongly annotated(i.e. available due to uploads by humans which have been manually tagged) and weakly annotated(i.e. available on blogs, newspaper reports, inaccurate tags by humans, etc). Typically the annotations are trained using specific images which have single annotation to them. However, this restricts the objects which can be identified to the dataset on which the automated annotater has been trained. We try to learn from commonly available images and their corresponding weak labels using a slight modification on the corr-LDA methodology.
Project supervised by Prof. Dinesh Babu Jayagopi.
Github: link