Research Projects

Artificial Intelligence (AI) has immense potential to transform healthcare diagnostics by enabling early, accurate, and scalable detection of medical conditions. However, a critical barrier to clinical adoption is the opacity of deep learning models, which often fail to provide human-interpretable reasoning essential for trust and accountability in healthcare. This project aims to advance foundational research to develop trustworthy, causally grounded AI systems for multi-modal medical diagnostics, building on our proven research in output-explainable deep models for blood smear image analysis and ECG analytics. Our team at the Indian Statistical Institute, with extensive expertise in machine learning, pattern recognition, and explainable AI, has developed systems aligning outputs with hematologist reasoning in blood smear analysis and domain knowledge overlays in ECG waveform interpretation. Leveraging these capabilities, we will design methods that systematically integrate diverse data modalities—including medical signals, images, and structured clinical information—within interpretable frameworks that clinicians can trust and adopt in workflows. We will specifically address key technical challenges including data heterogeneity, missing modalities, temporal and semantic misalignment across modalities, and the development of scalable, modality-robust encoders using pre-training on public datasets followed by fine-tuning on problem-specific datasets in collaboration with healthcare partners. We will design attention-based and co-attention transformer architectures for multi-modal fusion, explore domain-obedient self-supervision, and develop calibration and causal inference techniques to improve prediction reliability and actionable interpretability. Additionally, the project will develop advanced explanation modules using attention-based, prototype-based, and counterfactual techniques to provide clinician-aligned interpretability, and incorporate Vision-Language Model (VLM) based prompt-driven human-in-the-loop frameworks to enable clinicians to interact with and refine AI decisions dynamically. While the methods developed will have broad applicability across medical domains, we will demonstrate their impact in cardiology and hematology, leveraging ongoing collaborations with domain experts. In cardiology, we will integrate ECG and PPG signals, imaging, and clinical parameters for diagnostic and prognostic assessment, while in hematology, we will develop explainable models for blood smear image analysis to assist in identifying and characterizing blood disorders. These domains, chosen for their clinical significance and data richness, will serve as testbeds for validating our methods, ensuring that the developed AI systems align with real-world healthcare challenges and advance the science of explainable, reliable medical AI. This research aligns with national healthcare priorities, addressing the growing need for scalable, transparent, and effective diagnostic tools, particularly where specialist availability is limited. By enhancing trust in AI systems, the project will facilitate broader adoption of AI in healthcare, enabling professionals to deliver accurate, efficient, and explainable healthcare.