Bengaluru: Depression may not always announce itself as depression. It can show up as fatigue, a troubled stomach or simply a feeling that something is amiss. Now, an AI model being developed by researchers could help doctors detect those less obvious signs by analysing doctor-patient conversations — including the words, emotions and subtle cues that may otherwise go unnoticed.Called Human-in-the-loop Evaluation of Assisted Depression Screening (HEADS), the project launched here Thursday brings together Nimhans, IIT Kharagpur and Lokopriya Gopinath Bordoloi Regional Institute of Mental Health (LGBRIMH), Tezpur, to improve early detection of depression across languages and cultural contexts in India.The two-year project, running from Sept 2026 to Aug 2028, will develop and evaluate AI-assisted approaches in five languages — Kannada, Hindi, Bengali, Assamese and English. About 4,500 participants, including 4,000 patients and 500 volunteers from Nimhans and LGBRIMH, will be enrolled.The model will listen to conversations between physicians and patients and convert them into text. It will then translate conversations into English while retaining psychiatric terminology, idioms of distress and mixed-language speech — nuances that can often be lost during translation.Dr Lekhansh Shukla, assistant professor, Centre for Addiction Medicine, Nimhans, said depression is often missed because the first point of contact is usually not a psychiatrist trained to screen for and treat the condition. He said existing large language models remain weak at handling language and culture, especially Indian languages, where errors can be costly.Researchers first conducted a pilot study to benchmark existing AI systems. They found that even frontier models had a word error rate of 35% in Kannada clinical interviews, while only 64% of personal identifiers in clinical notes were correctly detected and removed.“Even on a relatively basic task of determining whether a phrase signifies a low mood, normal mood or a mood that is better than normal, most models perform quite poorly,” Shukla said.The system will remain a clinician-facing tool and will not be made directly available to the public, partly to prevent overdiagnosis.Nimhans director Dr Y C Janardhan Reddy said early identification of depression could also strengthen research and give greater confidence in diagnosis if differences are accounted for using AI models.Professor Animesh Mukherjee of IIT Kharagpur said the project will combine AI with speech-processing technologies to handle real-world clinical conversations. The team will also focus on safety and fairness, ensuring assessments are not unduly influenced by language, gender, educational or cultural background.——-BEYOND SYMPTOMS* HEADS will be developed in Kannada, Hindi, Bengali, Assamese and English* About 4,500 participants, including 4,000 patients and 500 volunteers, will be enrolled for two-year study* System will retain psychiatric terms, idioms of distress and mixed-language speech while translating conversations into English* Tool will remain clinician-facing, with the focus on accuracy, safety and fairness to avoid overdiagnosis
