Over the last few years, the digital innovation process and the COVID-19 pandemic spurred an increasing request for telehealth procedures [32, 33]. The first steps of the diagnostic process for ADHD may fit this trend, since a thorough information collection regarding children’s behaviors could be potentially performed remotely .
The first aim of the present study was to explore whether and to what extent the clinical diagnosis of ADHD by expert clinicians agreed with symptoms as rated by parents and teachers through online administered questionnaires. To this end, we tested a DT, given its notable interpretability and the suitability for digitally collected data .
Our algorithm reached a very good accuracy (82%) in correctly identifying children which either did or did not receive a diagnosis of ADHD at the end of the clinical evaluation. The present accuracy is in line with previous ML works which highlighted the possibility of accurately discriminating subjects with and without ADHD [13, 34, 35]. However, earlier research was based on biological, neurophysiological, or behavioral data collected on-site. To our knowledge, the present study provided first preliminary evidence that data collected through telehealth might be valuable to support the clinical practice of diagnosing ADHD.
As one could expect, among all the collected measures, the core parent- and teacher-reported ADHD symptom severity was the most discriminative information for the DT classification. Ratings on DSM-oriented ADHD scales of both the informants showed a crucial relevance for the clinicians’ diagnostic decision. This is interesting if we consider that the DT assigns the same “weight” to all the considered input variables (i.e., anamnestic, cognitive, behavioral). Moreover, although the algorithm was totally naïve about the questionnaire cutoffs, in the upper nodes the DT identified scores that are in line with moderate and severe risk for ADHD, respectively 64 and 70 [19, 23]. These findings thus extend –for the first time, in a telehealth setting– recent findings which showed that caregivers’ reports could reliably predict ADHD diagnosis .
Interestingly, in 18% of the cases, clinicians reached a different diagnostic conclusion compared to that resulting from the algorithm. Specifically, in case of “extreme” scores on caregivers’ reports, the clinicians rely mostly on both their direct observation of the patient, cognitive performance, and clinical interview for their decision.
The second aim of this study was to understand whether the co-presence of ASD symptoms as described by caregivers could represent a potential confounding factor, given the considerable overlap in symptom presentation . In our sample, 12% of children diagnosed with ADHD were also clinically diagnosed with ASD. This is in line with recent evidence .
It is important to keep in mind that the DT only relied on caregivers’ reports of ADHD core symptoms and often associated oppositional symptoms . As one could expect, ASD symptoms were not selected by the algorithm as discriminant information for a correct ADHD classification. Nevertheless, all participants clinically diagnosed with comorbid ADHD-ASD were correctly recognized as ADHD by the DT algorithm; conversely, not all participants with a clinical diagnosis of ADHD without an ASD diagnosis were correctly classified as ADHD. This therefore showed that both parents and teachers provided more severe ratings of ADHD in children who were later diagnosed with a comorbid ADHD-ASD by clinicians. The present finding is in line with a recent review, which reported higher externalizing difficulties in children with ASD  and with previous evidence describing an additive effect of symptom severity in children with ADHD/ASD comorbid state [39, 40].
Consistently, the analysis of social abilities among the four groups sorted by the DT showed that participants with higher ratings of social cognition, communication problems, and autistic mannerisms on SRS were classified as having ADHD, leading to many false positives for the algorithm. These traits, with the addition of social awareness, were conversely lower in children misclassified as non-ADHD by the algorithm. Despite representing a novelty for what concerns the analytic approach, this finding may corroborate previous evidence. Indeed, social functioning atypicalities, a hallmark of ASD, are often reported in ADHD, too . Although research suggests that the mechanisms underlying these difficulties are different , social impairment in the two conditions may look alike at phenotypic level to non-clinical observers such as parents. Hence, impaired social functioning can be reported by parents of children referred for suspected ADHD, influencing the DT results. An interesting exception to this trend is social motivation scores on SRS. All the four classes presented indeed typical levels of social motivation, which is in line with a recent work reporting comparable scores in social motivation between children with ADHD only and neurotypical peers .To our knowledge, the present study addressed for the first time the impact of ASD features on an ML algorithm classification of ADHD. Against the background of recent developments of digitalized procedures supporting diagnostic decisions, the confounding effect of non-core associated symptoms needs to be further investigated in future studies.
Online information collection and screening procedures should not be merely considered as an alternative to on-site diagnostic practice. Instead, telehealth can help effectively collect reliable caregivers’ reports and obtain a subsequent automated output regarding a diagnostic risk factor . Special attention should be given not only to the development of accurate diagnostic classification models but also to the factors that might lead to diagnostic misclassification. Lastly, if the first diagnostic steps are optimized, a reduction of the time delay between initial symptom detection and diagnosis could be achieved, enabling clinicians to focus on the most complex cases.
Some limitations of the present preliminary study should be considered. Our sample included children and adolescents from the same area (Northern Italy) referred by their pediatricians for suspected ADHD. Thus, it is not known whether our results could be generalized to other populations.
Furthermore, our analysis exclusively addressed the potentially confounding effects of autism symptoms in ADHD classification. However, there are several conditions commonly associated with ADHD . Additional research addressing the impact of these symptoms and conditions in predicting ADHD is needed.
4.3. Future directions
Future research focused on the development of online platforms for remotely performed data collection is needed . Developments of ML predictive models could also offer clinicians prompt feedback about the diagnostic risks associated with questionnaire scores. If proven valid, these procedures could be readily implemented also for other neurodevelopmental conditions.