Multinomial Logistic Regression for Interpretable Multiclass Decision Systems: A Systematic Review and Integrated Research Framework
Multinomial Logistic Regression (MLR) remains one of the most widely used interpretable models for multiclass classification, risk prediction, and discrete decision analysis. Its continued relevance does not reflect the novelty of the classical model, but the new demands placed on it by high-dimensional, noisy, imbalanced, and privacy-constrained data. This PRISMA-guided systematic review analyzes 108 peer-reviewed journal articles published or first made available online between January 2015 and December 2024, supported by a structured quantitative coding of the entire verified corpus. The bibliometric analyses and the methodological counts are both based on all 108 included studies. Four review questions examine the evolution of the literature, methodological extensions for modern data regimes, validation and reporting practices, and the derivation of an integrated research framework. The coded evidence reveals a method–validation–deployment gap: regularization and feature selection are common (feature selection in 48% of studies and penalized MLR in 29%), whereas the structured record documents class-specific metrics in 9% of studies, probability-calibration reporting in 1%, and code-availability reporting in 11%. The extraction fields do not support a defensible corpus-wide external-validation rate, and IIA or category-dependence assessment is documented in only 2% of the corpus. To make these findings actionable, we distinguish MLR-specific failure modes (IIA, separation, reference-category dependence, and calibration) from generic machine-learning concerns and propose an integrated framework spanning model scope, modeling approach, enabling technologies, and implementation and deployment, linking data regimes to suitable MLR strategies, validation requirements, and reporting expectations.