Using machine learning to improve diagnostic assessment of ASD in the light of specific differential diagnosis

Abstract

BackgroundDiagnostic assessment of ASD requires substantial clinical experience and is particular difficult in the context of other disorders with behavioral symptoms in the domain of social interaction and communication. Observation measures such as the Autism Diagnostic Observation Schedule (ADOS) do not take into account such comorbid and differential disorders.MethodWe used a well-characterized clinical sample of individuals (n=1262) that had received detailed outpatient evaluation for the presence of an ASD diagnosis (n=481) and covered a range of additional differential or overlapping diagnoses, including anxiety related disorders (ANX, n=100), ADHD (n=440), and conduct disorder (CD, n=192). We focused on ADOS module 3, covering the age range with particular high prevalence of such differential diagnoses. We used machine learning (ML) and trained random forest models on ADOS single item scores to predict a clinical best estimate diagnosis of ASD in the context of these differential diagnoses (ASD vs. ANX, ASD vs. ADHD, ASD vs. CD) and an unspecific model using all available data. We employed nested cross-validation for an unbiased estimate of classification performance (ASD vs. non-ASD).ResultsWe obtained very good overall sensitivity (0.89-0.94) and specificity (0.87-0.89) for the classification of ASD vs. non-ASD. In particular for individuals with less severe symptoms (around the ADOS cut-off) our models showed increases of up to 20% in sensitivity or specificity. Furthermore, we analyzed item importance profiles of the ANX-, ADHD- and CD-models in comparison to the unspecific model. These analyses revealed distinct patterns of importance for specific ADOS-items with respect to differential diagnoses.ConclusionUsing ML-based diagnostic classification may improve clinical decisions by utilizing the full range of information from comprehensive and detailed diagnostic observation such as the ADOS. Importantly, this strategy might be of particular relevance for individuals with less severe symptoms that typically present a very difficult decision for the clinician.

Abilities and Disabilities—Applying Machine Learning to Disentangle the Role of Intelligence in Diagnosing Autism Spectrum Disorders

“…Based on the present data, it can be suggested that going through the short DT pathways, to reinsure the decision for or against ASD, may decrease the risk of false positives and false negatives in individuals with below and above average IQ levels remarkably. However, it is relevant to keep in mind that, in addition to or in interaction with IQ, further factors may influence the complex diagnostic process, including other differential or comorbid diagnoses ( 17 , 18 , 36 , 45 ), female gender ( 46 ), parental psychiatric diagnoses ( 47 ), aspects of healthcare supply ( 48 , 49 ), and experience of the diagnostician ( 50 ). Further, since IQ and language are observed to account for the heterogeneity in ASD and the variability in diagnostic and therapeutically outcomes, it appears relevant to consider also developmental trajectories of autistic symptom severity and adaptive functioning ( 51 ).…”

Multi-axial Profile Scores (MAPS) for Complex Child Psychiatric Presentations

“…MAPS may be superior to screening tools as they are able to integrate information from multiple sources and weigh its significance for an individual. This is reflective of the clinical challenges in making sense of complex presentations where clinician experience and skill is required to integrate the data from assessment tools together with the clinical assessment in a nuanced manner to arrive at a diagnostic formulation which avoids diagnostic overshadowing and accounts for comorbidity and contributing bio-psychosocial context 13,14 .…”

“…This variability limits the clinical usefulness of categorical diagnostic models in complex presentations with multiple comorbidities [8][9][10][11] as their symptom profile may be non-specific to a particular diagnosis at the time of presentation. Additionally, psychometric tests may lead to false positives in complex cases, or, they also produce non-specific findings which are difficult to integrate into a formulation for an individual [12][13][14][15][16] . Experienced clinicians intuitively weight the data obtained from psychometric tools such as cognitive testing and autism assessment tools depending on the co-morbidities present.…”

“…Mismatch between demand and availability of child and adolescent psychiatrists creates space for new innovations to promote more equitable clinical care. Machine learning-based transdiagnostic models represent a potential method to reduce this inequity by earlier detection of neurodevelopmental disorders 12,13 . In order to be immediately clinically relevant, transdiagnostic models must capture the individualised weighting of differentiating characteristics which occurs intuitively in the experienced clinician 21 .…”

Conceptualization of the latent structure of autism: further evidence and discussion of dimensional and hybrid models

“…Furthermore, as the symptom profiles of the different classes demonstrate overlapping but also different profiles, the hybrid model appears more suitable for the early detection of ASD cases and may help to differentiate ASD from other mental disorders compared to a purely dimensional approach. Taking the symptom profiles into account could for example improve diagnostic accuracy [67].…”