KCDH Projects

A Machine learning based approach to identify proteomic networks and biomarkers driving Triple Negative Breast Cancer

• KCDH Faculty: Prof. Sanjeeva Srivastava, Bioscience and Bioengineering
• KCDH Partner: Prof. Sudeep Gupta, Department of Medical Oncology, Advanced Centre for Treatment Research and Education in Cancer, Tata Memorial Centre, Mumbai

Triple Negative breast cancer (TNBC) is the most aggressive subtype of breast cancer which accounts for 15-27% of all breast cancer cases in India, with the worst survival rate for any breast cancer subtype. TNBC is characterized by the lack of receptor targets which partly accounts for its poor prognosis and aggressive clinical manifestation. Herein we propose to utilize an ML based approach to identify proteomic expression networks that are responsible for tumorigenesis in TNBC. We will implement a machine learning and ANN-based Neural networking approach to develop an appropriate model from our discovery cohort and identified biomarkers will be validated in an independent set of patient samples.

Project Type: Sponsored by Koita Foundation

A novel Artificial Intelligence-based prognostic approach using PET-CT images and pathology images for advanced-stage Hodgkin Lymphoma

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital Health, IIT Bombay, Prof. Ganesh Ramakrishnan, Computer Science Engineering, IIT Bombay
• KCDH Partner: Dr. Hasmukh Jain, Medical Oncology, Tata Memoria Hospital, Mumbai

For advanced Hodgkin’s lymphoma, two chemotherapy regimens are often used for treatment, (i) ABVD and (ii) BEACOPP regimens. ABVD regimen has comparatively low cure rates as compared to the BEACOPP regimen, but it is tolerated well in most patients. BEACOPP, though associated with better outcomes, comes with the risk of significant toxicities. In this study, we propose to use AI/ML to evolve a strategy to help select the protocol based on the patient’s disease characteristics at baseline using PET-CT images and pathology images. This can help avoid the comparatively toxic BEACOPP protocol where ABVD can give equally good results. The study includes 1) a retrospective aspect, where we use the previously available datasets to develop AI models to predict the outcome of ABVD treatment, 2) the prospective aspect, where we will use the AI models to predict the response to the treatment of ABVD & BEACOPP chemotherapy protocols.

Project Type: Sponsored Project by Koita Foundation

An NMR-AI based Hybrid Approach for Modelling the Progression of Type 2 Diabetes Mellitus

• KCDH Faculty: Prof. Ashutosh Kumar, Biosciences and Bioengineering, Prof. Manjesh Hanawal, Department of Industrial Engineering and Operations Research
• KCDH Partner: Prof. Dr. Anurag Lila, Associate Professor, KEM Hospital and SGS Medical College, Mumbai

Type 2 Diabetes Mellitus (T2DM), a metabolic disorder - has become a primary global health concern. The world's second diabetic capital, India, is expected to witness a ~71.7% rise in diabetic cases by 2045. Currently used glucose level determination methods can report the disease long after the damage has been done and cannot be reversed. Therefore, in this proposal, we aim to decipher the metabolic patterns of individuals using high-throughput techniques like NMR and Mass Spectrometry. This data will then be subjected to different machine learning algorithms to generate a metabolic fingerprint for the early signature of diabetes. Then, fingerprints will be utilized to predict potential markers in ndividuals much before the onset of the disease. This personalized diagnosis will also be providing founding stones for personalized treatment. Diabetes is a metabolic disease, and metabolism is highly unique for every individual. Thus, personalized diagnosis, prevention, and diabetes treatment are needed.

Project Type: Sponsored by Koita Foundation

Assessing Quality of Cancer Care in India using Electronic Medical Records: A Pilot Study

• KCDH Faculty: Prof. Souvik Banerjee, Humanities and Social Sciences
• KCDH Partner: Dr. Sarbani Ghosh Laskar, Professor & Radiation Oncologist, Tata Memorial Hospital, Mumbai;

The overall goal of this study is to establish whether hospital-based electronic medical records (EMRs) can be used to measure quality of hospital care and to construct quality metrics that are feasible within existing EMR systems. To achieve this goal, we will (a) identify a core set of quality indicators, (b) assess whether and how they could be constructed with existing EMR data, (c) evaluate the practical weakly-supervised approaches needed to extract this data digitally from the EMR, (d) construct a subset of these indicators, and (e) provide recommendations on how to improve recording and use of data in EMR for monitoring quality of care. We will focus on oncology, which is the primary specialization of the study team. However, many of the quality indicators will be applicable to the general health care system. The project will be conducted within participating institutes that have implemented EMR systems over the past decade.

Project Type: Sponsored by Koita Foundation

Data Science Algorithms for Radiation Reduction in Computed Tomography

• KCDH Faculty: Prof. Ajit Rajwade, Computer Science Engineering
• KCDH Partner : Dr. Nitin Shetty, Associate Professor, Department of Interventional Radiology, Tata Memorial Centre, Mumbai; Mr. Nag Nath, Radiation Safety Officer (Scientific Assistant D), Tata Memorial Centre, Mumbai

The amount of radiation administered to patients, especially neonatal, geriatric or pregnant patients, and inadvertently to doctors and technicians, is a common concern in computed tomography (CT) scanning. The total radiation can be lowered by reduction in the number of tomographic projections acquired by a CT scanner (termed the "few views setting") and/or reduction in the XRay beam power per projection (termed the "low dosage setting". However this results in poorer quality image reconstruction, which could seriously interfere with accurate diagnosis. In many procedures, multiple CT scans per patient are acquired, either across weeks as in cancer surveillance, or in quick succession during CT-guided surgical procedures such as CT-guided biopsy or ablation. In these applications, the changes to the patient's anatomy across scans is restricted to relatively smaller areas. If the initial few scans are acquired with high radiation dosage, they can therefore act as guiding templates to enable more accurate reconstruction of the subsequent scans acquired from significantly low dosages. However while doing so, one must accurately determine the regions of genuine anatomical change across the scans, directly from the projection measurements. The aim of this project is to use machine learning algorithms to enable superior image reconstruction using the readily available prior information in such repeated scanning procedures. The long-term goal is to encourage CT scanner manufacturers to incorporate machine learning algorithms at the backend to enable more intelligent image scanning at lower dosages without sacrificing image quality.

Project Type: Sponsored by Koita Foundation

Designing predictive model for Childhood Congenital Glaucoma using Clinical and Genetic data

• KCDH Faculty: Prof. Abir De, Computer Science Engineering, Prof. Rishabh Iyer, Computer Science Engineering, UT Dallas
• KCDH Partner: Dr. B. Shantha, Sankara Nethralaya; Dr. Baskaran Mani, Sankara Nehralaya; Dr. S. Sripriya, Vision Research Foundation; Dr. Ronnie George, Vision Research Foundation

Childhood glaucomas account for a major form of blindness in children in developing nations, especially Primary Congenital Glaucoma (PCG, 1 in 10,000 live births in the West, 1 in 3300 live births in Andhra Pradesh as per Eye Diseases Study). We propose to devise predictive models using clinical and genetic data (n = 100) which will (a) primarily aim to identify genetic signals that can correlate with the clinical features (which can enable genetic counselling for expectant parents among surviving children in their 20s) (b) predict the surgical success and/or visual prognosis in children with PCG, and (c) exploratively predict the type of surgery that may be beneficial in PCG associated with specific genetic signals. This pilot project will lead to algorithms that can help clinicians counsel parents of children with blinding glaucoma for an overall better outcome.

Project Type: Sponsored by Koita Foundation

Discovery of novel biomarkers for the prediabetic stage with untargeted metabolomics and AI/ML guided data analysis.

• KCDH Faculty: Prof. Pramod P. Wangikar, Department of Chemical Engineering, IIT Bombay
• KCDH Partner: Dr. Rakesh Kumar Sahay, Endocrinology, Osmania Medical College, Hyderabad

Prediction of type 2 diabetes mellitus (T2DM) years before its onset would be valuable to the patient, their family, and the overall healthcare system of the country. These early warnings can be used to introduce lifestyle changes and manage the disease better. Prior reports have shown that certain cellular metabolites have a high predictive value for T2DM. For example, the plasma levels of branched-chain amino acids can predict T2DM up to 7-13 years before its onset1. However, all these studies have been reported in Western cohorts2. Our main objective is to identify predictive biomarkers in the Indian population. Since new biomarkers are to be identified from the hundreds of metabolites in human fluids, we propose to perform metabolomic profiling via high resolution (HR) liquid chromatography coupled to mass spectrometry (LCMS) and gas chromatography coupled to MS (GCMS). These techniques are well-known for their ability to perform untargeted and comprehensive analysis of a large number of metabolites simultaneously. Since these instruments generate large amounts of data per sample, key challenges include the separation of noise from true signals and annotation of unknown metabolite features. Thus, apart from acquisition of LCMS and GCMS data for human samples, our aim is to build AI-powered tools to automate the manual steps performed by the human experts including noise removal and spectral data annotation.

Wangikar’s group at IIT Bombay has expertise in metabolomics. We have developed advanced protocols for LCMS and GCMS-based identification and quantification of cellular metabolites. Our Collaborator Dr. Rakesh Sahay is an Endocrinologist who has over two decades of experience in Diabetes treatment and research. Dr. Sahay will be handling the clinical part of the project which includes (i) recruitment and classification of volunteers into different groups (T2DM, prediabetes, and healthy controls), (ii) collection of blood samples, (iii) running the approved diagnostic tests, (iv) shipment of samples to IIT Bombay, (v) clinical monitoring of volunteers for a prospective study and (vi) clinical correlation of the metabolomics results. IIT Bombay will handle the LCMS and GCMS data acquisition and data processing parts of the project. We have already completed the administrative and regulatory pre-requisites and we are ready to start the project immediately. This includes an MoU between IIT Bombay and Osmania Medical College and ethics committee approval from both the organizations (approval letters attached). Further, we have also analysed few human samples on LCMS and GCMS as a proof of concept for our study.

Project Type: Sponsored by Koita Foundation

Harmonized One-health Trans-Species Transmission of Antibiotics Resistance and levers for action using System Dynamics Modelling (HOTSTAR-SDM)

• KCDH Faculty: Dr. Om Damani, Department of Computer Science and Engineering
• KCDH Partner: Dr. Manoja Kumar Das, The INCLEN Trust International, New Delhi; Dr. Ashok Mahapatra, AIIMS Bhubaneswar; Dr. S. S. Mishra, Central Institute of Freshwater Aquaculture, Bhubaneswar; Dr. Basanti Pathi, Kalinga Institute of Medical Sciences, Bhubaneswar; Dr. Jayakrushna Das, Odisha University of Agriculture and Technology, Bhubaneswar

This proposal to be conducted at community level linked to hospitals in Odisha aims to document the environmental ‘Reservoirs of Resistance’, transmission dynamics and gradients of anthropogenic sources of antibiotic resistance (ABR) across the different hosts (humans, domesticated and food-chain animals/birds, fishes) and environment following one-health framework for antibiotics-resistant Escherichia Coli. The data and evidence base from multiple sources on antibiotics resistance ransmission dynamics and diffusion gradients for bacteria, antibiotic resistance genes (ARGs) and mobile genetic elements (MGEs) through multiple channels across the humansanimal / bird-fishes-environment and knowledge, attitude and practices (KAP) at household and community levels shall be used for the System dynamic modelling (SDM). The SDM application shall validate and refine the model causal loop diagrams (CLDs) prepared. The refined CLDs shall be tested for multiple hypothetical interventions for informing suitable levers for modification, policy and program level actions for preventing and reversing antibiotics resistance emergence and transmission.

Project Type: Sponsored by Koita Foundation

HMIS Analyzer: Visualization Dashboard for Health data

• KCDH Faculty: Prof. Sarthak Gaurav, Shailesh J. Mehta School of Management (SJMSOM), Prof. Satish Agnihotri (Co-PI) Emeritus Professor, Centre for Technology Alternatives for Rural Areas (CTARA)
• KCDH Partner: Rajalakshmi Nair, Nutrition Specialist, UNICEF State Office for Maharashtra

Health Management Information System (HMIS) is a web-based Monitoring Information System hosted by Ministry of Health & Family Welfare (MoHFW), Government of India. Around 200,000 health facilities across all districts of India upload data on the HMIS portal, each month. Data on 300 plus indicators are available in Excel and pdf format, and there is no customized data visualization tool available to aid interpretation of data and informed decision making by policymakers and programme managers. To address this need, we have developed a dashboard which currently has data on 25 key health indicators, up to district level for FY18-FY20, facilitating comparison across time, and amongst districts. This tool enables users to track progress of key indicators for performance coverage, and health outcomes. The proposed project will expand the set of indicators, deepen granularity, allow user-friendly and interactive visualisation on the dashboard with AI/ML based predictive capabilities among other analytics.

Project Type: Sponsored by Koita Foundation

Implementation of an alert system for early diagnosis of Acute Kidney Injury

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital health, Prof. Ganesh Ramakrishnan (Co-PI), Department of Computer Science and Engineering
• KCDH Partner: Dr. Barnali Das Kokilaben Dhirubhai Ambani Hospital & Medical Research Institute

Acute kidney injury (AKI) is a commonly occurring condition in patients admitted to the hospital associated with high rate of complications, morbidity and mortality. This disease is hard to diagnose owing to its asymptomatic pathogenesis and events of AKI are often missed by physicians. There is also a widespread unawareness regarding the dangers of this disease not only in patients but also physicians who are not nephrologists.

We intend to implement the first-ever alert system, for early diagnosis of AKI, in India. In the pre-implementation phase of the study, we will create an algorithm that will track changes in serum creatinine in all eligible patients, as per the Kidney Disease Improving Global Outcomes (KDIGO) criteria. In the post-implementation phase, we will evaluate if there are improvements in patient-related outcomes as well as clinician behavior in comparison to the pre-implementation phase findings. This study will help us identify the epidemiological trends of AKI patients in an Indian hospital setting, evaluate the potential of an alert system in diagnosing AKI early leading to early intervention and possibly mitigating the poor outcomes associated with this silent killer. This study will also set a benchmark for public health advisories in an Indian healthcare setting for preventing the adversities of an easily preventable disease.

Project Type: Sponsored by Koita Foundation

To develop Machine learning-based models and web-based tools to predict the mild, moderate and severe cases of Covid 19 based on routine laboratory biomarkers.

• KCDH Faculty: Prof. Biplab Banerjee, Centre of Studies in Resources Engineering (CSRE) & Centre for Machine Intelligence and Data Science (C-MInDS)
• KCDH Partner: Dr. Barnali Das, Biochemistry & Immunology Division, Kokilaben Dhirubhai Ambani hospital & Medical Research Institute.

COVID patients as mild, moderate, and severe classes based on the analysis of the biomarkers using advanced deep learning techniques. Specifically, we are interested in the scenarios where less amount of annotated data are available and data can be noisy. We plan to look into the aspects of few-shot learning to deal with the scenario. Further, we are interested in performing a study on the importance of the biomarkers in properly classifying the disease scenarios. Finally, we propose to develop a web-based tool to continually learn the model, and annotate data globally by the users.

Project Type: Sponsored by Koita Foundation

Use of Wearable sensors and Pranayama-based interventions for improving mental-wellbeing of Undergraduate Students at IITB

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital Health(KCDH), IIT Bombay,
• KCDH Partner: Dr. Nirmal Punjabi, Sensorama Technologies LLP

Mental wellbeing of undergraduate students is significantly affected during the transition to campus system. Majority of these cases go unnoticed owing to the lack of effective diagnostic systems. Recent development in wearable systems now enable us to monitor the mental wellbeing on a continuous and long-term basis. This project will leverage the new wearable sensor technology to capture various digital biomarkers (heart rate variability, breathing rate etc.) along with machine-learning to enable holistic monitoring of mental wellbeing of the student population. The project will also establish the effectiveness of active intervention techniques such as pranayama and meditation to improve the mental wellbeing. Project focus will be to measure and establish the mental wellbeing score using the wearable sensors technology and validate the findings with clinical biomarkers (cortisol, epinephrine, and norepinephrine levels). Pranayama and other breathing exercises effectiveness in improving the mental wellness will be established on >100 students using digital and clinical biomarkers.

Project Type: Sponsored by Koita Foundation

Addressing Class Imbalance in Medical Image Data: Finding Needle in a Haystack

• Supervisor: Prof. Ganesh Ramakrishnan
• Co-supervisor: Prof. Rishabh Iyer

Current state-of-the-art active learning algorithms (www.arxiv.org/abs/1906.03671, www.arxiv.org/abs/2012.10630, www.cse.iitb.ac.in) do not often work well in cases where there is a class imbalance (e.g. the cancerous cases are much lower compared to the non-cancerous cases), distribution shift (e.g. training with data from one ethnicity/group and testing on another ethnicity), and out of distribution examples (e.g. unseen classes in the unlabeled set). All of these issues are often present in real-world medical imaging datasets. We will develop new active learning algorithms, which can effectively address the issues pointed out above. We will achieve this by building upon the supervisors’ work on active learning using submodular mutual information functions www.arxiv.org/abs/2103.00128. To get a glimpse of the datasets and challenges in this space, look at www.grand-challenge.org. As for the coding platform, we will develop on www.decile.org - our homegrown data efficient machine learning platform. More specifically, we will build on the DISTIL (www.github.com/decile-team/distil) - an open source platform for Deep diversified interactive Learning with several Jupyter notebooks and video tutorials (such as www.youtube.com).

The overall goal of the project is Data and Cost Efficient Deep Learning for Medical Image Classification. The advancement of machine learning and deep learning is creating a big impact in several domains. One such important domain is to build machine learning classifiers to effectively complement human doctors and radiologists in detecting various diseases (e.g. forms of cancer) from various medical images, e.g. X-rays, CT scans, and MRI images. Examples of applications range from cancer tumor detection, medical image segmentation, to detection of Alzheimer’s and Parkinson’s disease. Despite all the amazing progress of deep learning, one main challenge of these approaches is that deep learning models are extremely data-hungry and require several tens of thousands of images to work effectively. Given that detecting diseases like cancers in images requires specialized skill-sets (doctors and radiologists), the cost of annotating these datasets is very high. Furthermore, it is often hard to find sufficient samples for certain rare diseases, and medical imaging data is often heavily class-imbalanced. The goal of this project is to significantly reduce the amount of labeled data required, with minimal loss in accuracy. To achieve this, we will study the role of active learning and semi-supervised learning to reduce the amount of labeled data required. Semi-supervised approaches aim at effectively using the unlabeled data in complementing the limited labeled data in learning, while data selection and active learning seek to select the most informative labeled data to improve the model performance. Preliminary results suggest that we can reduce the amount of labeled data by factors of 5x to 20x with negligible performance degradation, depending on the dataset and the choice of the algorithm.

The advisors complement each other in this project: Ganesh Ramakrishnan's and Rishabh Iyer’s expertise is in active learning and semi-supervised learning and Prof. Tamil's expertise is in deep learning for medical imaging. We will facilitate a summer internship at a Healthcare technology company/organization, as part of the collaboration, subject to approval by the faculty advisors.

Project Type: Student Project

Advancing causal inference methods in health economics and health services research

• Supervisor: Prof. Souvik Banerjee

The proposed study will explore use of latent factor methods to obtain causal estimates of treatment effects and its applications in healthcare settings. I have already published work on this topic and would like to extend the econometric framework.

Project Type: Student Project

Automated feature selection for biomarker discovery from big biological data

• Supervisor: Prof. Pramod Wangikar
• Co-supervisor: Prof. Ganesh Ramakrishnan

Proteins and metabolites, the new class of biomarkers are expected to bring a paradigm shift in the diagnosis, monitoring and treatment of human disease and will make personalized medicine a reality in near future. Moreover, the next generation biomarkers are likely to be based on the inference drawn from multiple metabolite or protein molecules rather than single measurements such as the blood glucose level that is currently used for the diagnosis of diabetes. The present project focuses on the discovery of biomarkers from big biological data involving genomics, proteomics or metabolomics. Our primary objective will be to select the best combination of biomarkers or minimum subset of features to predict class labels such as disease vs. healthy or one category of disease vs. another.

The challenges include:

We will use a number of feature selection methods and machine learning tools together with the concepts of multitask learning to achieve the task of biomarker discovery. A large number of public domain databases are available that will be used as test cases. See a recently published paper to understand the broad objectives of the proposed project (www.frontiersin.org). You may also see literature on multitask machine learning. with special focus on data efficient machine learning (see www.decile.org). The supervisor is active in the fields of metabolomics, systems biology and big biological data analysis. The candidate must have a strong foundation in the mathematics and programming that enable AI/ML, apart from the willingness to work in cross-disciplinary areas. We will try and facilitate a summer internship at a Healthcare technology company/organization, subject to approval by the faculty advisors.

Project Type: Student Project

Deep active learning for medical image classification

• Supervisor: Prof. Ganesh Ramakrishnan
• Co-supervisor: Prof. Rishabh Iyer

We will develop new active learning algorithms for the specific domain of medical imaging, which can effectively combine aspects of uncertainty (picking examples to label where the current model is most confused on) and diversity (picking representative and non-redundant examples to label) to pick the most informative samples for the specific tasks at hand. Furthermore, we will also study the sample efficiency and labeling cost reduction which can be achieved by active learning. To get a glimpse of the datasets and challenges in this space, look at www.grand-challenge.org. As for the coding platform, we will develop on www.decile.org - our homegrown data efficient machine learning platform. More specifically, we will build on the DISTIL ( www.github.com/decile-team/distil) - an open source platform for Deep diversified interactive Learning with several Jupyter notebooks and video tutorials (such as www.youtube.com).

The overall goal of the project is Data and Cost Efficient Deep Learning for Medical Image Classification. The advancement of machine learning and deep learning is creating a big impact in several domains. One such important domain is to build machine learning classifiers to effectively complement human doctors and radiologists in detecting various diseases (e.g. forms of cancer) from various medical images, e.g. X-rays, CT scans, and MRI images. Examples of applications range from cancer tumor detection, medical image segmentation, to detection of Alzheimer’s and Parkinson’s disease. Despite all the amazing progress of deep learning, one main challenge of these approaches is that deep learning models are extremely data-hungry and require several tens of thousands of images to work effectively. Given that detecting diseases like cancers in images requires specialized skill-sets (doctors and radiologists), the cost of annotating these datasets is very high. Furthermore, it is often hard to find sufficient samples for certain rare diseases, and medical imaging data is often heavily class-imbalanced. The goal of this project is to significantly reduce the amount of labeled data required, with minimal loss in accuracy. To achieve this, we will study the role of active learning and semi-supervised learning to reduce the amount of labeled data required. Semi-supervised approaches aim at effectively using the unlabeled data in complementing the limited labeled data in learning, while data selection and active learning seek to select the most informative labeled data to improve the model performance. Preliminary results suggest that we can reduce the amount of labeled data by factors of 5x to 20x with negligible performance degradation, depending on the dataset and the choice of the algorithm.

The advisors complement each other in this project: Ganesh Ramakrishnan's and Rishabh Iyer’s expertise is in active learning and semi-supervised learning and Prof. Tamil's expertise is in deep learning for medical imaging. We will facilitate a summer internship at a Healthcare technology company/organization, as part of the collaboration, subject to approval by the faculty advisors.

Project Type: Student Project

Disparities in access to and utilization of healthcare services in India

• Supervisor: Prof. Souvik Banerjee

This study will examine disparities in access to and utilization of healthcare services in terms of geography (rural vs. urban), religion, and socio-economic status using nationally-representative dataset in India.

Project Type: Student Project

Implementation of an alert system for early diagnosis of Acute Kidney Injury

• Supervisor: Prof. Siuli Mukhopadhyay
• Co-supervisor: Dr. Barnali Das

Acute kidney injury (AKI) is a commonly occurring condition in patients admitted to the hospital associated with high rate of complications, morbidity and mortality. This disease is hard to diagnose owing to its asymptomatic pathogenesis and events of AKI are often missed by physicians. There is also a widespread unawareness regarding the dangers of this disease not only in patients but also physicians who are not nephrologists.

We intend to implement the first-ever alert system, for early diagnosis of AKI, in India. In the pre-implementation phase of the study, we will create an algorithm that will track changes in serum creatinine in all eligible patients, as per the Kidney Disease Improving Global Outcomes (KDIGO) criteria. In the post-implementation phase, we will evaluate if there are improvements in patient-related outcomes as well as clinician behavior in comparison to the pre-implementation phase findings. This study will help us identify the epidemiological trends of AKI patients in an Indian hospital setting, evaluate the potential of an alert system in diagnosing AKI early leading to early intervention and possibly mitigating the poor outcomes associated with this silent killer. This study will also set a benchmark for public health advisories in an Indian healthcare setting for preventing the adversities of an easily preventable disease.

Project Type: Student Project

Semi supervised learning in Medical image data

• Supervisor: Prof. Ganesh Ramakrishnan
• Co-supervisor: Prof. Rishabh Iyer

A large amount of unlabeled data that are available in the medical imaging arena can be put to use through semi-supervision learning (SSL), where the unlabeled data is effectively used in the training procedure. Through the use of the unlabeled data, semi-supervised learning has shown accuracy improvements of around 3 - 5%, compared to supervised learning with the small labeled set (www.arxiv.org/abs/2001.07685, www.arxiv.org/abs/2008.09887). We can additionally use active learning to select the most informative samples thereby reducing the amount of labeled data required to achieve the same accuracies in SSL. To get a glimpse of the datasets and challenges in this space, look at www.grand-challenge.org. As for the coding platform, we will develop onwww.decile.org - our homegrown data efficient machine learning platform. More specifically, we might build on the SPEAR ( www.github.com/decile-team/spear) - an open source platform for Semi-supervised data programming with several Jupiter notebooks and video tutorials.

The overall goal of the project is Data and Cost Efficient Deep Learning for Medical Image Classification. The advancement of machine learning and deep learning is creating a big impact in several domains. One such important domain is to build machine learning classifiers to effectively complement human doctors and radiologists in detecting various diseases (e.g. forms of cancer) from various medical images, e.g. X-rays, CT scans, and MRI images. Examples of applications range from cancer tumor detection, medical image segmentation, to detection of Alzheimer’s and Parkinson’s disease. Despite all the amazing progress of deep learning, one main challenge of these approaches is that deep learning models are extremely data-hungry and require several tens of thousands of images to work effectively. Given that detecting diseases like cancers in images requires specialized skill-sets (doctors and radiologists), the cost of annotating these datasets is very high. Furthermore, it is often hard to find sufficient samples for certain rare diseases, and medical imaging data is often heavily class-imbalanced. The goal of this project is to significantly reduce the amount of unlabeled data required, with minimal loss in accuracy. To achieve this, we will study the role of active learning and semi-supervised learning to reduce the amount of labelled data required. Semi-supervised approaches aim at effectively using the unlabelled data in complementing the limited labelled data in learning, while data selection and active learning seek to select the most informative labelled data to improve the model performance. Preliminary results suggest that we can reduce the amount of labelled data by factors of 5x to 20x with negligible performance degradation, depending on the dataset and the choice of the algorithm.

The advisors complement each other in this project: Ganesh Ramakrishnan's and Rishabh Iyer’s expertise is in active learning and semi-supervised learning and Prof. Tamil's expertise is in deep learning for medical imaging. We will facilitate a summer internship at a Healthcare technology company/organization, as part of the collaboration, subject to approval by the faculty advisors.

Project Type: Student Project

Addressing intensive care unit drudgery with Large Language models in low resource settings

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital Health, IIT Bombay
• KCDH Partner: 10BedICU, India

The nature of the work in an intensive care unit (ICU) can be intense and often involves life-or-death situations, which necessitates a structured and disciplined environment where strict protocols are followed for standardized patient care and round the clock operations putting an enormous burden on healthcare professionals. 

Recently, we developed a large language model (LLM) on the top of a pretrained GPT 3.5 to address the repetitive questions raised by onsite medical professionals. This LLM based AI assistant (named Ayushma) was set to use only the inputted corpus as a source for responding to prompts without using the pretrained knowledge base to reduce the incidence of ‘hallucinations’. A series of real-life ICU questions were used as prompts to Ayushma and compared to freely available GPT3.5 and GPT4 LLMs which were assessed by a panel of expert doctors. We demonstrate that with the guard-rails of a highly curated corpus and not using the pretrained knowledge-base, Ayushma generated highly concordant responses (judged by panel of experts) compared to other LLMs. This demonstrates a highly useful use-case of LLMs in the medical setting for low resource settings to address the drudgery of ICUs. 

Project Type: Faculty Research Project undertaken at KCDH

Large language Models for India Centric Healthcare Support

• KCDH Faculty: Prof. Ganesh Ramakrishnan, Computer Science Engineering, IIT Bombay, Prof. Kshitij Jadhav, Koita Centre for Digital Health, IIT Bombay
• KCDH Partner: Dr. Vishnu Vardhan, Vizzhy Inc., Vardhan SK Healthcare, Hyderabad

The proposal aims to explore the challenges and opportunities of implementing large AI models in low- and middle-income countries (LMICs) like India. By creating a national data pool, researchers can access diverse datasets to develop India-centric AI solutions, improve services, address disparities, and promote social equity. The project will use Optical Character Recognition (OCR), Automatic Speech Recognition (ASR), and machine translation to build a large corpus for the proposed LLM. The implementation of this endeavor could lead to significant individual and national benefits, fostering development, collaboration, and data-driven policies.

Project Type: Sponsored by Vardhan SK Healthcare, Hyderabad (Vizzhy Inc)

Implementation of Electronic Alert System for Acute Kidney Injury in Hospital Setting and Development of Web Based Tools for National Public Health Advisory in India.

• KCDH Faculty: Prof. Siuli Mukhopadhyay, Department of Math
• KCDH Partner: Dr. Barnali Das, Consultant, Biochemistry & Immunology Division, Laboratory Medicine, Kokilaben Dhirubhai Ambani hospital & Medical Research Institute.

Acute kidney injury (AKI) is a commonly occurring condition in patients admitted to the hospital associated with high rate of complications, morbidity and mortality.

Our aim is to identify patients at risk for AKI in the admitted patients to track the rise in serum creatinine value and detect stages of AKI cases as per the Kidney Disease Improving Global Outcomes (KDIGO) staging criteria. This would enable preemptive evaluation and intervention to minimize risk and AKI severity. An AKI risk prediction algorithm and a web-based tool as per the KDIGO criteria, using electronic health record data patients will be built.

After AKI diagnosis, we will evaluate if there are improvements in patient-related outcomes and dialysis. This study will also help us identify the epidemiological trends of AKI patients in an Indian setting, evaluate the potential of an alert system in diagnosing AKI early leading to early intervention and possibly mitigating the poor outcomes associated with this. It will also set a benchmark for public health advisories in an Indian healthcare setting for preventing the adversities of an easily preventable disease.

The search for novel biomarkers of acute kidney injury, Urinary NGAL is one of the earliest and most highly induced proteins of kidney damage. We will also perform one small pilot study to evaluate the role of urinary NGAL in all the clinically validated cases of AKI.

Project Type: Sponsored by Koita Foundation

A complete OCR framework for historical medical records digitization

• KCDH Faculty: Prof. Ganesh Ramakrishan, Computer Science & Engineering Department, IIT Bombay
• KCDH Partner: Dr. Devi Shetty, Narayana Health

Over the past few decades, in the process of the IT revolution, many hospitals adopted Hospital Information & Medical Records Software(HIMS) to help them store medical records in electronic format. Most of the time, the HIMS were complementary to physical medical records. This led to most of the Electronic Medical Records being stored as scanned images. Though the documents are always available for the doctors to view, the fact that the documents can be viewed only as an attachment has its disadvantages such as the inability to search or group the records based on the content. A point in case is the medical partner has a large corpus of the scanned medical records. The goal of this project is to ensemble a set of tools, some of which will use cutting-edge AI technologies, to create a workflow for the digitization of historic medical records.

Project Type: Sponsored by Koita Foundation

HIerarchical FEDERAted Learning for the INdian healthcare SYSTem (HIFEDERAL- INSYST)

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital Health(KCDH), IIT Bombay
• KCDH Partner: Trust Lab, IIT Bombay

The implementation of Machine Learning (ML) in the healthcare industry is facing difficulties due to the decentralized nature of healthcare data, which is stored in silos within hospitals. However, a new method called Federated Learning (FL) has emerged as a promising solution. FL enables a central entity to gather insights from decentralized data sources without sharing the raw data.

We propose a partially decentralized design of FL. Our research will focus on resolving the following key questions: 1) how to design a hierarchical communication topology and aggregation algorithm across different data distributions, 2) how to reason about different levels of privacy and convergence across different levels of hierarchy, and 3) rigorously evaluating the effectiveness of our proposed FL algorithm and other federated learning algorithms on healthcare datasets within a hierarchical topology.

Project Type: Students Sponsored by Trust Lab, IIT Bombay

A study to develop an artificial intelligence algorithm for predicting the outcomes of febrile neutropenia (FN) in adolescent and adults with haematological malignancies

• KCDH Faculty: Prof. Amit Sethi, Department of Electrical Engineering, IIT Bombay,
• KCDH Partner: Dr. Hasmukh Jain, Medical Oncology, Tata Memorial Hospital, Mumbai

Febrile Neutropenia (FN) is one of the most frequent and serious complications of chemotherapy. A wide spectrum of pathogens is known to cause FN. A specific microbial agent is, however, isolated only in 23% of the cases 1. Delay in initiating appropriate antibacterial therapy results in high rates of mortality 3 and morbidity. IDSA and ECIL guidelines developed in the west help in the selection of antibiotics at the beginning and during the course of FN. These guidelines are largely empirical which leads to the use of an excess of antibiotics Artificial intelligence can be employed to predict the risk of complications and select the choice of antimicrobials. It can be done by utilizing the high number of data available from medical records and optimizing the big data to train with deep learning methods.

We propose to conduct a study to develop an artificial intelligence algorithm to predict the risk of FN and select antibiotics in cancer patients receiving chemotherapy.

Project Type: Sponsored by Koita Foundation

Developing an ML algorithm to identify precancerous lesions from oral mucosal images

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital Health(KCDH), IIT Bombay
• KCDH Partner: Oriacle Bioscience Ltd

Tobacco use is a significant risk factor for oral mucosal lesions, including oral pre-cancer and cancer, particularly in India and parts of Asia. The availability of commercial tobacco preparations has increased in these regions, contributing to the rise in cancerous and precancerous lesions in the oral cavity. Early detection of oral cancer is crucial for better survival rates, but challenges like lack of awareness and limited healthcare resources hinder early identification. To address this, we aim to develop classification algorithms using Medical Data Efficient Machine Learning (Med-DECILE) to identify oral precancerous lesions. We plan to use shallow ML models as labeling functions on limited labeled data and pseudo-label unlabeled data through Data Programming, improving accuracy iteratively in a semi-supervised learning setting.

Project Type: Sponsored by Oriacle Bioscience Ltd

Multimodal Data Analytics for Identification of Multi-Omic Biomarkers of Diabetes Mellitus

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital Health(KCDH), IIT Bombay
• KCDH Partner: Dr. Vishnu Vardhan, Vardhan S K Healthcare (Vizzhy, Inc)

Type 2 diabetes (T2D) is a metabolic disorder with high blood sugar levels. Its management includes dietary changes, physical activity, and medications like sulfonylureas and insulin. T2D is a global health concern, and India alone may have 70 million cases by 2025. At KCDH, we aim to develop new prevention strategies by studying the association between T2D and gut microbiome using large datasets from the Human Microbiome Project. We'll also explore how diet and drugs can influence microbiome changes.

Project Type: Sponsored by Vardhan S K Healthcare Inc (Vizzhy, Inc)

Semi automatic segmentation and classification of CT/MR images: through ischemic stroke and sarcopenia as case studies

• KCDH Faculty: Prof.Vikram Gadre, Department of Electrical Engineering, IIT Bombay,
• KCDH Partner: Prof. Rama Jayasundar, AIIMS New Delhi

CT and MR (functional and diffusion) are both used to capture images of internal body organs. They are the best non-invasive ways to help doctors with diagnosing certain diseases, an otherwise difficult task. However, in most cases, the CT/MR images are interpreted by radiologists. The process is laborious and error-prone. Artificial Intelligence (AI) and Deep Learning (DL) techniques can help doctors with the process of studying and evaluating disease/treatment using the MR/CT images along with patient history. Based on retrospective studies, in this proposal, we use AI/DL techniques to study medical images (CT/MR) along with other patient histories to study two medical problems: (i) Sarcopenia, (ii) Stroke. With primary focus on Stroke and based on time period availability sarcopenia will also be studied. We will use Active Learning methods to help reduce the radiologist effort in labelling medical imaging data and employ weak and semi-supervised approaches that can effectively leverage a small amount of labelled data in conjunction with a large unlabelled dataset.

In this study, we will evaluate AI-based deep learning techniques for diagnostic and clinical outcome markers through retrospective analysis of single time point MR/CT images for fat and sarcopenia indices and time series MR images for rehabilitation of stroke patients.

Project Type: Sponsored by Koita Foundation

Investigating host and gut microbiome interactions in Diabetes mellitus

• KCDH Faculty: Prof. Kshitij Jadhav, Koita Centre for Digital Health(KCDH), IIT Bombay
• KCDH Partner: Wadhwani AI, New Delhi, India

Type 2 diabetes (T2DM) is characterized by high blood sugar levels due to insulin resistance and impaired secretion. The gut microbiome plays a role in T2DM, influencing glucose metabolism and inflammation. Studies on enterotypes reveal distinct clusters of gut microorganisms linked to different health outcomes. The Human Microbiome Project explores the gut's impact on health using advanced sequencing techniques. The aim is to reevaluate the gut microbiome's role in diabetic development by considering specific enterotypes, which may enhance the significance of less common microbes. We can reanalyze longitudinal metagenome data from prediabetic individuals to understand enterotype trends during diabetic progression and explore temporal dynamics. The newly generated enterotype information will also assist in assessing the functional contributions of less abundant microbiomes in diabetic metabolic imbalances.

Project Type: Sponsored by Wadhwani AI