Health Sathi AI

Hospitals produce a huge amount of free-text material every day, from prescriptions to discharge notes, yet very little of it can be searched or analysed in a structured way. Medical concept recognition, which finds clinical mentions in text and ties them to standard identifiers, is the usual route to unlock it. This paper reviews three families of approaches (dictionary and rule-based systems, the MedCAT toolkit, and transformer-based models) and then examines one practical case in detail: adapting a pre-trained MedCAT model to prescription-style text containing Indian medicine names, carried out during an internship at C-DAC Noida. Instead of relying on retraining alone, the work concentrated on the model’s concept database, adding Indian brand names, spelling variants and synonyms, and fine-tuning on 500 synthetic records. On a small in-house test set, synonym mapping raised symptom recognition from 40.00% to 86.36%, disease recognition from 61.54% to 86.54% and medicine recognition from 21.84% to 43.68%. A coverage check also showed that the baseline model matched only 229 of 1,808 Indian medicines (12.67%). The review concludes that the quality of the concept vocabulary deserves as much attention as the choice of model, and it lists the evaluation gaps that still need to be closed.