Anirudh Gangadharan

I build diagnostic systems for settings where they don't exist. My research focuses on cross-modal medical inference - extracting clinical insight from low-cost inputs (smartphone images, ECG waveforms) to replace expensive diagnostics in resource-constrained environments. Current work includes LLM-based ECG-to-echocardiography inference and a smartphone hemoglobin estimation pipeline validated in 600 pregnant women. I am a third-year medical student (MBBS) at Gadag Institute of Medical Sciences. My long-term direction is building translation layers between biological systems and computational interfaces.

Research & Innovation

ECG is All You Need [Clinical Study — n=311]

Co-Investigator

Demonstrated first proof-of-concept for cross-modal cardiac inference: LLM-generated echocardiographic reports from 12-lead ECG images alone, without fine-tuning.
Retrospective analysis of 311 paired ECG–echocardiogram cases comparing three LLMs against cardiologist gold standard.
Best model achieved r=0.394 (p<0.001) for ejection fraction prediction, with 64% of estimates within ±5% of true values.
Aimed at democratizing cardiac imaging in resource-constrained settings where echocardiography is unavailable.

Screening for Left Ventricular Dysfunction from Photographed 12-Lead ECGs [Deep Learning — n=500]

Co-Investigator

Trained a multi-task EfficientNet-B3 ensemble to predict ejection fraction and regional wall-motion abnormality directly from photographed 12-lead ECGs.
Three-stage transfer learning (ImageNet → PTB-XL 21,430 ECGs → 500 paired records) with inverse-density loss weighting and out-of-fold cross-validation.
Severe-EF detection AUROC 0.76, RWMA screen AUROC 0.70; best folds reach 0.83–0.86. Diagnosed and fixed shortcut-learning and regression-to-the-mean failure modes via Grad-CAM.
Deployed as a multilingual WhatsApp triage tool with a public Hugging Face demo.

HemoglobinAI [Clinical Study – n=600]

Principal Investigator

Developed a multimodal machine learning pipeline for non-invasive hemoglobin estimation using smartphone imaging.
Compared palpebral conjunctiva and buccal mucosa images in 600 pregnant women using standardized color calibration.
Implemented YOLOv8-based segmentation and evaluated seven regression models with 5-fold cross-validation.
Demonstrated marginally superior predictive performance using palpebral conjunctiva imaging.

EyeScopeAI [ICMR STS Grant]

Principal Investigator

Developing a fully offline computer vision system for refractive error screening in rural settings.
Secured competitive national research funding (ICMR STS) to address healthcare access disparities through frugal innovation.

LumiSleep [Neurostimulation]

Founder & Architect

Conceived a closed-loop neurostimulation system using sensory entrainment; pre-incubated at Manipal Bioincubator.
Architected proprietary algorithms for nervous system regulation and sleep optimization.

DoctorsAI

Advisory Board Member

Leading national scaling of student wings and mentoring interdisciplinary projects bridging clinical needs with data science.
Co-organized Global Summit 2025; strategic guidance for AI literacy programs for 5,000+ medical professionals.

Selected Publications

AI in Cardiovascular Medicine: Wearables & Real-time Monitoring.
AI in Medicine: Clinical Updates & Physician Perspectives (Jaypee Publications, 2025). ISBN: 9789372020472.
Future Directions & Innovations in AI for Medicine.
AI in Medicine: Clinical Updates & Physician Perspectives (Jaypee Publications, 2025). Primary Author.
ECG is All You Need: Democratising Cardiac Imaging with LLMs.
Doctors AI Global Summit 2025 Scientific Abstract Compendium. McGraw Hill Education.
Screening for Left Ventricular Dysfunction from Photographed 12-Lead ECGs: A Deep-Learning Feasibility Study.
National Conference on AI for Doctors, Solapur 2026.