Application of Neural Networks and ANFIS in System Identification for Blood Glucose Prediction

This project explores the use of neural networks and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) for system identification of the Bergman glucose-insulin model, aiming to provide a new data-driven method for blood glucose prediction in diabetic patients. This study combines control theory, biomedical engineering, and artificial intelligence, retaining the interpretability of classical models while enhancing data-driven prediction capabilities, which is of great significance for precision medicine and personalized health management.

Diabetes is a global public health challenge. Accurate blood glucose prediction is crucial for insulin adjustment and hypoglycemia prevention. Traditional management relies on empirical judgment and struggles to capture complex dynamics. The Bergman model is a classic glucose-insulin kinetic model with physiological interpretability, but its parameter estimation is complex and it is difficult to adapt to individual differences and time-varying characteristics, so it urgently needs improvement by combining modern machine learning.

System identification establishes dynamic system models through input-output data. The blood glucose system is nonlinear, time-delayed, and subject to random disturbances. This project uses two methods: neural networks (strong function approximation ability) and ANFIS (combining the interpretability of fuzzy logic with the adaptive learning of neural networks) to address the shortcomings of traditional linear methods.

Targeting the time-series characteristics of blood glucose, recurrent neural networks (RNN) and LSTM are used to capture long-term dependencies. Input features include historical blood glucose, insulin injections, carbohydrate intake, and time encoding. Training uses time-series cross-validation to ensure generalization, and the loss function considers clinical relevance (e.g., higher weight for hypoglycemia prediction).

ANFIS transforms fuzzy inference into a multi-layer network, which can integrate expert knowledge (e.g., rising blood glucose + low insulin → high blood glucose risk) and optimize rules through data. Input variables include blood glucose change rate, insulin action intensity, postprandial time window, etc. The membership function balances complexity and overfitting, providing stronger interpretability than pure neural networks (NN).

Data comes from public continuous glucose monitoring datasets. Preprocessing includes outlier removal (sensor failure), physiological constraint interpolation to fill missing values, and normalization. Feature engineering extracts multi-scale time features (short-term trends, medium-term postprandial responses, long-term rhythms) and explores the comparison between personalized and general models.

In addition to RMSE/MAE, the evaluation focuses on clinical indicators (hypoglycemia prediction sensitivity, hyperglycemia warning accuracy). Results show that both NN and ANFIS outperform the traditional Bergman model. ANFIS is suitable for scenarios requiring manual review, NN is better at complex nonlinearity, and their integration has a synergistic effect.

Application prospects include closed-loop control of artificial pancreas, mobile health warning, and clinical decision support. Translation challenges: need to verify robustness in large-scale populations, pass medical AI regulatory approval, improve user trust, and solve data privacy and security issues. Interdisciplinary integration provides new ideas for biomedical problems.