# Development and Internal Evaluation of an Interpretable AI-Based Composite Score for Psychosocial and Behavioral Screening in Dental Clinics Using a Mamdani Fuzzy Inference System

## TL;DR

This paper introduces a new interpretable AI system to combine three psychosocial and behavioral scores into a single dental screening tool, which is robust and coherent.

## Contribution

A novel deterministic Mamdani fuzzy inference system is developed to integrate GAD-7, PHQ-9, and OBC-21 into a unified psychobehavioral composite score for dental screening.

## Key findings

- The composite score (PCS) showed strong correlations with individual measures (GAD-7, PHQ-9, OBC-21) and increased with symptom severity.
- PCS demonstrated excellent robustness to input perturbations and membership function variations.
- Concordance with linear baselines was high, with minor nonlinearity observed at higher scores.

## Abstract

Background and Objectives: Psychosocial symptoms and oral behaviors can complicate routine dental care, yet available screeners yield multiple separate scores. Explainable artificial intelligence offers a pragmatic way to integrate such multidomain measures into a single, auditable output that can support screening-oriented stratification and standardized documentation (non-diagnostic). Therefore, we aimed to develop an interpretable, deterministic Mamdani fuzzy inference system (FIS) integrating GAD-7, PHQ-9, and OBC-21 into a 0–10 psychobehavioral composite score (PCS) to support screening-oriented stratification and standardized documentation (non-diagnostic). Materials and Methods: Cross-sectional multicenter study in 18 private dental clinics in Romania (October 2024–March 2025; n = 460). A rule-based Mamdani Type-1 FIS was specified a priori (48 rules; triangular membership functions; centroid defuzzification) without supervised training. Internal evaluation assessed coherence across severity strata, robustness to predefined input perturbations (±1 point; ±5%) and membership-function variation (±10%), and benchmarking against linear composites (Z-mean; PCA PC1). Results: Median PCS was 2.30 (IQR 2.03–3.56). PCS correlated with GAD-7 (Spearman ρ = 0.886), PHQ-9 (ρ = 0.792), and OBC-21 (ρ = 0.687) (all p < 0.001), increased monotonically across anxiety and depression severity strata, and was higher in high OBC-21 risk. Robustness was excellent under input perturbations (ICC(3,1) = 0.983 for ±1 point; 0.992 for ±5%) and high under ±10% membership-function variation (ICC(3,1) = 0.959). Concordance with linear baselines was high (Spearman ρ = 0.956 for Z-mean; 0.955 for PCA PC1), with a small systematic nonlinearity at higher scores. Conclusions: PCS provides a fully auditable, rule-based integration of three patient-reported measures with coherent internal behavior and robustness to plausible measurement noise and specification changes. This study reports internal evaluation of a deterministic, rule-based aggregation; external clinical validation against independent outcomes is required before any clinical utility claims.