John Snow Labs 57d ago

Clinical trials optimise for control. Real patients do not behave that way. The real signal of treatment effectiveness is when patients actually switch. Timestamps 00:00 Why clinical trials fail to reflect real oncology outcomes 01:52 What TTNT measures and why it matters in practice 03:54 Dataset and feature design from real-world oncology data 06:03 Survival-informed ML approach (not full survival modelling) 08:42 Model performance and why feature quality dominates 10:23 Feature importance and clinical interpretability 11:44 Real-world applications and deployment impact Time-to-Next-Treatment (TTNT) reframes oncology analytics from controlled endpoints to real-world behaviour. It captures when therapy stops working in practice, not just in trials. Hemant Dandu (Associate Director, Data Science & Advanced Analytics at IQVIA) presents a survival-informed ML approach built on longitudinal patient data, where treatment transitions become a measurable, operational signal rather than a retrospective insight. The model is intentionally simplified to a binary classification problem, but the signal design carries survival logic through feature engineering. Treatment history, progression indicators, and utilisation patterns dominate performance, not model complexity. Multiple models converge to similar results. That is the point. Signal quality, not algorithm choice, determines reliability in this setting. The outcome is directly actionable: identifying patients likely to transition early enables proactive intervention, segmentation, and treatment planning across clinical and commercial teams. 📌 Applied Healthcare AI Summit 2026 — real-world AI in healthcare, from pilots to production systems. #OncologyAI #HealthcareAI #RealWorldData #MachineLearning #ClinicalAI #PredictiveAnalytics