Context-Aware adaptive normalization LSTM (CAAN-LSTM) for immunotherapy decision support in cancer clinical data analysis.

Abstract

Clinical decision-making for cancer immunotherapy is challenged by the heterogeneous and often incomplete nature of patient time-series data. Traditional models struggle to account for individual patient variability and frequently underperform in real-world settings due to static normalization techniques and limited handling of missing data. We introduce CAAN-LSTM, a novel deep learning architecture that dynamically adjusts its normalization strategy during sequence processing by incorporating patient-specific clinical context. The key components of the model include: A meta-learned adaptive normalization layer, attention mechanisms for fusing temporal and static data, a hypernetwork that generates personalized scaling and shifting parameters, a transformer-based context encoder to model static patient features, a learned masking strategy to manage missing values and quantization-aware training to enable efficient deployment across diverse hardware platforms. In terms of prediction accuracy, CAAN-LSTM performed noticeably better than traditional models, particularly in situations with high variability or insufficient data. The model showed strong potential for assisting with individualized cancer treatment planning after being successfully piloted in Saudi medical facilities and validated on real-world clinical datasets. CAAN-LSTM offers a robust and adaptive framework for modeling clinical time-series data. By integrating patient-specific context and real-time normalization, it enhances decision support for immunotherapy and is well-suited for practical clinical implementation.