Repositório Institucional

Repositório de Publicações Científicas

Preservar, Divulgar e Dar Acesso à Produção Intelectual

DA UNIVERSIDADE PORTUCALENSE

 
Publicações
Pessoas
Projetos
Organizações

Entradas recentes

  • Item type: Publicação , Access status: Acesso aberto ,
    A Differentiable Keras-Based Adaptive Neuro-Fuzzy Inference System for Diabetes Prediction Using NHANES 1999–2018 Data
    (2026-10-21) Narne, Sravanthi; Vemuri, Srikanth; Jammalamadaka, Sastry Kodanda Rama; Jammalamadaka, Sasi Bhanu; Thottempudi, Pardhu; Moreira, Fernando
    Diabetes mellitus is a major public health concern, creating a need for reliable computational approaches for diabetes status classification and risk-oriented clinical assessment. Conventional machine learning models can capture predictive relationships in clinical data but may offer limited interpretability, while purely fuzzy approaches may have limited learning capability and scalability. This study proposes a differentiable Keras-based Adaptive Neuro-Fuzzy Inference System (ANFIS) for diabetes prediction using clinically relevant variables from the National Health and Nutrition Examination Survey (NHANES) 1999–2018 dataset. The proposed framework integrates trainable Gaussian membership functions, fuzzy-rule-based inference, and neural consequent learning within a unified TensorFlow–Keras computational graph. The fuzzy and neural parameters are jointly optimised via gradient-based learning with the Adam optimiser and binary cross-entropy loss. The experimental dataset contains 18,085 samples and uses glucose, blood pressure, body mass index (BMI), and age as input features. On the independent test set, the proposed model achieved 92.54% accuracy, 71.66% precision, 62.04% sensitivity, 96.67% specificity, 66.50% F1-score, 92.57% ROC-AUC, and 79.35% balanced accuracy. The results show that the proposed framework combines nonlinear predictive learning with structured fuzzy reasoning. The Keras-based implementation further provides a flexible foundation for integration with modern deep learning and healthcare analytics pipelines. Nevertheless, further validation using larger, heterogeneous, and multi-institutional clinical datasets is required before clinical deployment.