Table 2.

Summary of recent and representative studies aiming to reduce the number of behavioral measures needed for ASD diagnosis through multivariate analysis. Reported sample sizes are the numbers used for classification and do not necessarily reflect the study’s total sample size.

Reference	Study Participants	Experimental Methods	Key Features	Multivariate Technique	Key Results
Wall et al. (2012)176	891 ASD and 75 non-spectrum children for training; 1976 ASD and 1000 simulated controls for validation	Used ADI-R data from three ASD data repositories	7 of 93 items from the ADI-R	Alternating decision tree, chosen from 15 algorithms	99.9% training accuracy, with both 99.9% sensitivity and specificity when predicting the validation set
Wall et al. (2012)177	612 ASD and 15 non-spectrum children for training; 446 ASD and 1000 simulated controls for validation	Used ADOS data from three ASD data repositories; pilot study for Duda et al. (2014)178	8 of 29 items from ADOS Module 1	Alternating decision tree, chosen from 16 algorithms	100% training accuracy, with 99.7% sensitivity and 94% specificity when predicting the validation set
Duda et al. (2014)178	2333 ASD and 283 non-spectrum children	Used ADOS data from five ASD data repositories	ADOS Module 1 feature set from Wall et al. (2012)177	Alternating decision tree	Validated with 98% sensitivity and 77% specificity against the original ADOS
Wilson et al. (2014)179	58 male adults with ASD and 66 TD controls	Administered three ASD evaluations and nine neuropsychological tests/tasks	Ten variables from performed tasks, plus verbal IQ and performance IQ	SVM	Achieved 81% accuracy, 78% sensitivity, and 85% specificity with leave-two-out cross-validation
Kosmicki et al. (2015)180	362 (510) ASD and 282 (93) non-spectrum individuals for training; 1089 (1924) ASD and 66 (214) non-spectrum for validation†	Used data from five ASD data repositories; evaluated score sheets separately for ADOS Module 2 and Module 3	9 of 28 behaviors from Module 2 and 12 of 28 behaviors from Module 3	Logistic regression (Module 2); radial kernel SVM (Module 3)	99% sensitivity and 89% specificity for Module 2 validation; 98% sensitivity and 97% specificity for Module 3 validation
Bone et al. (2016)181	1264 verbal individuals with ASD and 462 verbal individuals without ASD	ADI-R and Social Responsiveness Scale items taken from a data repository	Five behavioral codes total from the two assessments	SVM	Classified individuals below (above) age 10 with 89% (87%) sensitivity and 59% (53%) specificity
Cohen et al. (2016)182	535 children with ASD and 125 children without ASD	PDD Behavior Inventory forms collected from five sites	Six domain scores of PDD Behavior Inventory, parent-reported	CART	82%/83%/86% sensitivity and 88%/87%/93% specificity for training/testing/validation
Levy et al. (2017)183	1319 (2870) ASD and 70 (273) non-ASD children†	ADOS Module 2 and Module 3 score sheets from four ASD data repositories	Nine items from Module 2 and nine from Module 3	Logistic regression; SVM	Classified with 89%/95% sensitivity and 90%/87% specificity for Module 2/3
Feczko et al. (2018)184	47 children with ASD and 58 TD children	Had children perform seven tasks related to information processing	34 behavioral variables related to performed tasks	Random forest	Achieved 73% classification accuracy, 63% sensitivity, and 81% specificity

^†Numbers outside (inside) parentheses indicate sample sizes for analyzing ADOS Module 2 (Module 3).