Currently, ADHD is diagnosed by a clinical expert, typically a psychiatrist or specialist paediatrician. However, clinical assessment takes a minimum of 1 hour, and there is a global shortage of trained specialists, which means that diagnosis is often delayed (Adesman, 2001). Therefore, it is important to consider alternative approaches to improve the efficiency of early diagnosis, including the role of artificial intelligence. In this study, available literature on machine learning and deep learning studies on ADHD diagnosis were reviewed, and the various diagnostic tools used were identified.

In this study, five databases were searched for relevant publications up to December 2021, and 92 of the retrieved articles were found to be eligible for inclusion in this review. The following types of ADHD diagnostic tools utilised to develop artificial intelligence models were identified:

• Brain magnetic resonance imaging (MRI) was identified as the most widely studied modality for automated ADHD diagnosis, with 39/92 studies analysing brain MRI images of patients with ADHD. More studies had implemented machine learning (28/39), compared with deep learning (11/39), in the MRI analysis for ADHD. Brain functional connectivity was identified as the most common input feature for ADHD diagnosis. Most of the studies obtained their MRI images from the Neuro Bureau ADHD-200 Preprocessed Repository (ADHD-200) public database (Bellec et al, 2017).
• Physiological signals were used to detect ADHD in 24/30 studies (electroencephalogram [EEG]: 23/39; electrocardiogram: 1/39). Studies using physiological signals for the detection of ADHD had high model performances (all >80% for machine learning and deep learning models). Power spectral features were the most commonly extracted type of feature from EEG signals.
• Questionnaires/rating scales for the identification of ADHD were used in 6/30 studies, and only machine learning models were proposed. Decision tree classifiers, including random forest classifiers, were found to be commonly used to analyse questionnaire data.
• Game simulation and performance tests were used in two machine learning studies each to train their models to identify ADHD.
• Motion data can be used in the diagnosis of ADHD and were used in 4/39 studies. These data demonstrated increased activity level as a feature of ADHD. It was noted that the accelerometer device is advantageous over EEG, because it is an unobtrusive measurement technique, allowing the recording of the patient in their natural environment.
• Miscellaneous data were the least common modality of ADHD diagnosis used in machine learning studies; 2/39 studies used pupillometric data, and only one study used Twitter data.

The authors highlighted the increase in research regarding ADHD diagnosis over the years, in particular the deep learning models. Furthermore, multiple validation methods employed by the machine learning and deep learning studies were identified, the most prevalent methods being hold-out validation and 10-fold cross-validation.

It was acknowledged that, due to the lack of publicly available ADHD data, the majority of the studies in this review used private datasets. Furthermore, the number of study participants, the data collection procedures and the validation methods varied greatly across studies, making it difficult to directly compare results across studies.

The authors concluded that the majority of machine learning and deep learning studies on ADHD diagnosis retrieved for this review utilised MRI and EEG; other modalities were reported by very few studies. Additionally, the lack of publicly available datasets for the majority of modalities, except for MRI, was acknowledged. The authors recommended that future research should focus on developing more publicly available datasets for other modalities in ADHD assessment, as well as developing artificial intelligence models that utilise data from wearable devices for ADHD diagnosis and monitoring. The authors also suggested that future artificial intelligence studies in ADHD should improve the interpretability of their models to facilitate their adoption in a clinical setting.

Read more about automated detection of ADHD here

Disclaimer: The views expressed here are the views of the author(s) and not those of Takeda.