An Interpretable Min-Max Layered Neural Network for Real-Time Immune Feedback Monitoring in Immunotherapy

Cytokines, immune cells, and immune checkpoint molecules are interconnected and create feedback loops that either activate or suppress immunity. Correct modelling of these dynamic interactions is critical to understand immune regulation and to develop adaptive immunotherapeutic approaches. Traditional deep learning models, on the other hand, tend to be black-box predictors, which lack biological interpretability and clinical transparency. This paper presents a new Min-Max Layered Neural Network (MML-NN) architecture to model and monitor immune feedback processes in real-time. The architecture includes biologically inspired min-max operations to model threshold-dependent immune activation, inhibition signals and temporal feedback, and increase the interpretability. The framework was tested through simulated longitudinal immune biomarker data such as cytokine profiles, CD8⁺ T-cell counts and PD-L1 expression, and is compared to Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), feedforward neural network and Transformer-based models as proof of concept. Robustness was evaluated by 10 independent experimental runs, 95% confidence interval analysis and by component-wise ablation studies. The full MML-NN attained a 93.10% prediction accuracy, 91.40% F1 score, six time-step prediction lead time and computational latency of 12.5 ms, while surpassing the comparison models in terms of predictive accuracy, early immune-state transition detection, and interpretability. Repeated experiments showed low variability, while ablation results confirmed the importance of all architectural components. The framework was validated using simulated data, but it offers a foundation for future clinical validation and AI-driven precision immunotherapy decision support.