Establishing diagnostic thresholds for adult ADHD screening in taxi drivers: validation of the CAARS-S via Item Response Theory and Machine Learning.
L’essentiel
Adult Attention-Deficit/Hyperactivity Disorder (ADHD) is under-recognized in professional drivers, yet it poses significant safety risks. The Conners' Adult ADHD Rating Scale Short Version (CAARS-S:SV) lacks empirically validated cutoff scores for occupational screening in Iran. This study aimed to assess the psychometric properties of the Persian CAARS-S:SV in taxi drivers and to determine optimal ADHD diagnostic thresholds using both traditional and machine-learning methods. A sample of 298 male taxi drivers (mean age 36.8 ± 8.9 years) completed the Persian CAARS-S:SV. Internal consistency was evaluated with Cronbach's α, and construct validity was tested via confirmatory factor analysis (CFA). The dataset was randomly split into training (n = 198) and test (n = 100) subsets. Diagnostic thresholds were derived using receiver operating characteristic (ROC) curve analysis, item response theory (IRT; Terluin et al. method), logistic regression, and Random Forest (RF) modeling. Performance metrics-sensitivity, specificity, positive predictive value, and negative predictive value-were computed on the test set. Cronbach's α coefficients ranged from 0.72 (Subscale A) to 0.89 (total score), and CFA fit indices supported the three-factor structure (χ2/df = 2.40, RMSEA = 0.07; CFI = 0.91). ADHD prevalence was 20.8% (n = 62). ROC and RF methods identified optimal total-score cutoffs yielding sensitivity = 88% and specificity = 86.7%. For Subscale D, ROC achieved sensitivity = 96% and specificity = 76%, while RF achieved sensitivity = 92% and specificity = 85.3%. However, RF resulted in lower error value to diagnose the status of ADHD in the participants. IRT, ROC, and RF matched for subscales A, B, and C but produced an unacceptably low sensitivity (68%) for the total score (threshold = 24), reflecting a violation of IRT's unidimensionality assumption. Logistic regression performed less favorably (sensitivity 16%-60%). The Persian CAARS-S:SV is a reliable and valid instrument for ADHD screening in Iranian taxi drivers. ROC curve and Random Forest methods yielded the most accurate cutoff scores, particularly for the overall scale and dimension D. These empirically derived thresholds can enhance early identification of adult ADHD in occupational health settings and support the development of data-driven screening protocols.
Synthèse détaillée
Résumé original
Adult Attention-Deficit/Hyperactivity Disorder (ADHD) is under-recognized in professional drivers, yet it poses significant safety risks. The Conners' Adult ADHD Rating Scale Short Version (CAARS-S:SV) lacks empirically validated cutoff scores for occupational screening in Iran. This study aimed to assess the psychometric properties of the Persian CAARS-S:SV in taxi drivers and to determine optimal ADHD diagnostic thresholds using both traditional and machine-learning methods. A sample of 298 male taxi drivers (mean age 36.8 ± 8.9 years) completed the Persian CAARS-S:SV. Internal consistency was evaluated with Cronbach's α, and construct validity was tested via confirmatory factor analysis (CFA). The dataset was randomly split into training (n = 198) and test (n = 100) subsets. Diagnostic thresholds were derived using receiver operating characteristic (ROC) curve analysis, item response theory (IRT; Terluin et al. method), logistic regression, and Random Forest (RF) modeling. Performance metrics-sensitivity, specificity, positive predictive value, and negative predictive value-were computed on the test set. Cronbach's α coefficients ranged from 0.72 (Subscale A) to 0.89 (total score), and CFA fit indices supported the three-factor structure (χ2/df = 2.40, RMSEA = 0.07; CFI = 0.91). ADHD prevalence was 20.8% (n = 62). ROC and RF methods identified optimal total-score cutoffs yielding sensitivity = 88% and specificity = 86.7%. For Subscale D, ROC achieved sensitivity = 96% and specificity = 76%, while RF achieved sensitivity = 92% and specificity = 85.3%. However, RF resulted in lower error value to diagnose the status of ADHD in the participants. IRT, ROC, and RF matched for subscales A, B, and C but produced an unacceptably low sensitivity (68%) for the total score (threshold = 24), reflecting a violation of IRT's unidimensionality assumption. Logistic regression performed less favorably (sensitivity 16%-60%). The Persian CAARS-S:SV is a reliable and valid instrument for ADHD screening in Iranian taxi drivers. ROC curve and Random Forest methods yielded the most accurate cutoff scores, particularly for the overall scale and dimension D. These empirically derived thresholds can enhance early identification of adult ADHD in occupational health settings and support the development of data-driven screening protocols.