Manuscript under review at Heliyon

In this research, I developed an LLM-based diagnostic chatbotpowered by GPT-4o for early diagnosis of 14 diseases that are prevalent in South Asia. The system was designed to emulate the structured reasoning process of a clinician. Additionally, to ensure transparency and enhance the accuracy of the system, Feature Attribution, Chain of Thought prompting and a two-phase diagnostic system were introduced.

1. Symptom-Based Diagnosis: The system asks the patient follow-up questions based on their condition and updates the most likely disease set after each patient response. Finally, it ranks the top three most likely diseases and moves on to phase 2.
2. Test Result-Based Diagnosis: In this phase, the model requests relevant laboratory test values for these three candidates and produces the final diagnostic output.

The model demonstrated promising performance, acquiring an accuracy of 90% and top-3 accuracy of 100%, which is supported by local and global symptom matching using cosine similarity. The hyperparameters were chosen through an elaborate ablation study, and fine-tuning strategies were implemented to mitigate hallucinations in the LLM.

Ongoing Research — Oct 2025 – Present

Developing a causal counterfactual segmentation framework to improve causal consistency and disease-progression explainability, trained and evaluated on the BRSET dataset (16,266 images from 8,524 patients).