Weighted Probabilistic ADHD Models integrate dimensional and categorical approaches to ADHD assessment and Improve Diagnostic Discrimination Beyond DSM-5 Symptom count

Technical Summary: Weighted Probabilistic ADHD Models

Problem Statement
Current diagnostic frameworks for Attention-Deficit/Hyperactivity Disorder (ADHD), specifically the DSM-5-TR, rely on a categorical symptom-count approach. This method assigns equal diagnostic weight to all symptoms within the inattention and hyperactivity–impulsivity domains, regardless of their individual predictive validity, developmental stage, or sex-specific expression. Evidence suggests that this "equal weighting" assumption is flawed: symptoms vary in their discriminatory value, and strict cutoffs (e.g., 6 out of 9 symptoms for children, 5 out of 9 for adults) may exclude individuals with significant clinical impairment or bias diagnosis toward male presentations. Furthermore, purely categorical systems fail to capture the continuous, dimensional nature of psychopathology, potentially limiting the precision of individualized assessment required for modern psychiatry.

Methodology
The study utilized a retrospective observational design with a clinical database of 584 patients aged 6–22 years referred for ADHD assessment. Participants were categorized based on clinician-confirmed DSM-5 diagnoses (reference standard) into ADHD-positive (n=499) and ADHD-negative (n=85) groups. Exclusion criteria included Autism Spectrum Disorder, psychotic disorders, IQ <70, or pending diagnoses.

The researchers developed and compared three multivariable logistic regression models against the conventional unweighted DSM-5 symptom-count model:

1. Weighted Adjusted Model (Model 1): Incorporates symptom-specific coefficients adjusted for sex and categorical age group (>12 vs. ≤12 years).
2. Symptom-by-Age Interaction Model (Model 2): Adds interaction terms between symptoms and age to test if developmental stage modifies symptom predictive value.
3. Symptom-by-Sex Interaction Model (Model 3): Adds interaction terms between symptoms and sex to test for sex-specific predictive weighting.

Model performance was evaluated using Receiver Operating Characteristic (ROC) analysis, exact DeLong testing for pairwise comparisons, calibration plots, Brier scores, and Decision Curve Analysis (DCA). Internal validation was conducted via 1,000 bootstrap resamples to estimate optimism-corrected discrimination. Optimal probability thresholds were determined using the Youden index.

Key Results

• Diagnostic Discrimination: The weighted adjusted model (Model 1) significantly improved diagnostic discrimination compared to the unweighted DSM-5 symptom count for both dimensions:
  • Hyperactivity–Impulsivity: AUC increased from 0.973 (DSM-5) to 0.976 (Weighted Model); DeLong p = 0.013.
  • Inattention: AUC increased from 0.913 (DSM-5) to 0.937 (Weighted Model); DeLong p = 0.007.
• Symptom Heterogeneity: Regression coefficients revealed substantial variation in the diagnostic weight of individual symptoms. For example, in the hyperactivity domain, "Acting as if driven by a motor" (β=2.08) and "Blurting out answers" (β=2.12) carried significantly higher weights than "Interrupting or intruding" (β=1.08). In the inattention domain, "Fails to pay attention to details" (β=1.92) was weighted higher than "Appears not to listen" (β=0.70, p=0.077).
• Interaction Models: More complex models incorporating symptom-by-age or symptom-by-sex interactions (Models 2 and 3) did not significantly improve discrimination beyond the parsimonious Weighted Adjusted Model (Model 1). This suggests that while individual symptoms may vary by age or sex, incorporating these interactions as main effects in the weighted model is sufficient without adding structural complexity.
• Optimal Thresholds: The Youden index identified optimal probability thresholds of 0.600 for hyperactivity–impulsivity (Sensitivity 93.4%, Specificity 95.7%) and 0.689 for inattention (Sensitivity 96.3%, Specificity 85.0%).
• Calibration and Utility: Weighted models demonstrated lower Brier scores (indicating better prediction accuracy) and higher net benefit across clinically relevant threshold probabilities in Decision Curve Analysis compared to DSM-5 approaches. Internal validation confirmed minimal overfitting (optimism estimates of 0.001 and 0.002).
• Subgroup Performance: While performance was robust across most strata, the inattention dimension showed lower discrimination specifically in females older than 12 years (Youden index = 0.592), highlighting the limitations of categorical thresholds in this subgroup.

Significance and Claims
The paper claims that weighted probabilistic models offer a practical, clinically interpretable framework that bridges dimensional and categorical approaches to ADHD assessment. By quantifying the differential diagnostic value of specific symptoms and adjusting for age and sex, these models improve diagnostic discrimination beyond traditional symptom counting without sacrificing usability.

The authors emphasize that unlike "black-box" machine learning techniques, logistic regression-based weighting preserves transparency, allowing clinicians to understand the specific contribution of each symptom. The proposed approach allows for the generation of individualized probability estimates while still providing binary classification decisions via empirically derived thresholds, facilitating integration into electronic health records or clinical calculators.

The study concludes that while the weighted models significantly outperform DSM-5 symptom counts, the added complexity of interaction models is unnecessary for improving discrimination. The authors note limitations, including the reliance on a single-center retrospective dataset and the use of clinician diagnosis (which itself is DSM-based) as the reference standard. They suggest future work should focus on external multicenter validation and prospective evaluation of treatment response.