Publication: Hybrid method mechanism of multi-layer feed-forward neural network (MLFFNN) for biometry regression model
No Thumbnail Available
Date
2026-01
Authors
Adnan, Mohamad Nasarudin
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Statistical modeling plays a critical role in medical and forensic sciences,
where reliable prediction of complex biological and investigative data is essential.
Traditional regression models have contributed significantly but often fail to capture
nonlinear relationships and higher-dimensional patterns that arise in modern datasets.
Despite advances in statistical and machine learning techniques, existing approaches
remain limited in predictive accuracy and robustness, particularly when applied to
challenging biometry data such as chronic disease prediction and forensic analysis. A
methodological gap persists in integrating conventional regression with machine
learning to create a unified, reliable, and interpretable modeling framework. This study
aims to develop and elucidate a hybrid methodological mechanism that integrates
Ordered Logistic Regression (OLR), Multiple Linear Regression (MLR), and Binary
Logistic Regression (BLR) with Multi-Layer Feed-Forward Neural Networks
(MLFFNN), Response Surface Plot (RSP), and Decision Tree Analysis (DTA) for
biometry modeling. A hybrid model was constructed by combining classical
regression techniques with MLFFNN for nonlinear pattern detection, enhanced by
extensive bootstrapping for robust parameter estimation. Data were preprocessed to
remove outliers, resampled 1000 times to stabilize estimates, and divided into training
and testing sets. OLR was applied for ordinal outcomes, MLR for continuous
outcomes, and BLR for binary outcomes, with each regression output further
optimized using MLFFNN by tuning hidden layers, activation functions, and learning
rates. RSP was used to visualize multi-variable interactions and determine optimal Statistical modeling plays a critical role in medical and forensic sciences,
where reliable prediction of complex biological and investigative data is essential.
Traditional regression models have contributed significantly but often fail to capture
nonlinear relationships and higher-dimensional patterns that arise in modern datasets.
Despite advances in statistical and machine learning techniques, existing approaches
remain limited in predictive accuracy and robustness, particularly when applied to
challenging biometry data such as chronic disease prediction and forensic analysis. A
methodological gap persists in integrating conventional regression with machine
learning to create a unified, reliable, and interpretable modeling framework. This study
aims to develop and elucidate a hybrid methodological mechanism that integrates
Ordered Logistic Regression (OLR), Multiple Linear Regression (MLR), and Binary
Logistic Regression (BLR) with Multi-Layer Feed-Forward Neural Networks
(MLFFNN), Response Surface Plot (RSP), and Decision Tree Analysis (DTA) for
biometry modeling. A hybrid model was constructed by combining classical
regression techniques with MLFFNN for nonlinear pattern detection, enhanced by
extensive bootstrapping for robust parameter estimation. Data were preprocessed to
remove outliers, resampled 1000 times to stabilize estimates, and divided into training
and testing sets. OLR was applied for ordinal outcomes, MLR for continuous
outcomes, and BLR for binary outcomes, with each regression output further
optimized using MLFFNN by tuning hidden layers, activation functions, and learning
rates. RSP was used to visualize multi-variable interactions and determine optimal
predictor levels, while DTA generated interpretable, rule-based decision paths. Model
performance was assessed using accuracy, sensitivity, specificity, Mean Squared Error
(MSE), and Mean Absolute Deviance (MAD), and benchmarked against singlemethod
models. The proposed hybrid framework consistently outperformed
conventional techniques across three applied case studies: (i) type 2 diabetes among
dyslipidemia patients, (ii) forensic crime case, and (iii) hypertension among
dyslipidemia patients. It achieved higher predictive accuracy, significantly reduced
MSE and MAD, and provided improved interpretability through DTA and RSP
visualizations. The findings demonstrate the model’s ability to capture complex
nonlinear patterns, strengthen diagnostic reliability, and enhance predictive
performance across both medical and forensic domains. This research introduces a
validated hybrid statistical–machine learning mechanism that advances biometry
regression modeling, offering methodological innovation and practical tools for
healthcare diagnostics and forensic investigations.