Post by Samantha Weber, PhD

🚀 Introducing 𝑴𝑨𝑫𝑹𝑺-𝑩𝑬𝑹𝑻, featured in our latest publication “Using a fine-tuned large language model for symptom-based depression evaluation” - published in 𝑛𝑝𝑗 𝐷𝑖𝑔𝑖𝑡𝑎𝑙 𝑀𝑒𝑑𝑖𝑐𝑖𝑛𝑒. ▶️ https://lnkd.in/eaFp5c6M In this work, we fine-tuned a lightweight German BERT model to predict individual depressive symptom scores from structured clinical interviews. 🔍 𝗞𝗲𝘆 𝗳𝗶𝗻𝗱𝗶𝗻𝗴𝘀 • Achieved 79–88% accuracy (on a 7-point clinical scale) and mean average error (MAE) 0.7–1.0, performing at clinician level. • Fine-tuning reduced prediction errors by 75% vs the base model. • Scales efficiently - strong performance even with limited data - and can be readily deployed in low-resource clinical settings. 💡 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 Depression is still diagnosed largely through subjective assessments. Our results show that small, fine-tuned LLMs - not massive models - can deliver accurate, interpretable, and scalable tools for clinical monitoring and symptom assessment in mental health care. 💡💻 𝗙𝘂𝗹𝗹𝘆 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲: code, synthetic data, and model weights → 🤗 https://lnkd.in/ecAdfw_d Kudos to all the co-authors and collaborators - Nicolas Deperrois, Robert Heun, Laura Frühschütz, Anna Monn (née Bankwitz), Stephanie Homan, PhD, Andrea Häfliger, Tobias Kowatsch, Birgit Kleim, Sebastian Olbrich and the entire MULTICAST consortium. University of Zurich Psychiatrische Universitätsklinik Zürich Centre for Digital Health Interventions School of Medicine - Universität St.Gallen (HSG) Swiss National Science Foundation SNSF #AI #MentalHealth #DigitalPsychiatry #LLM #Depression #Evaluation #BERT #MADRS