Telehealth Autism Assessment Passes a Key Reliability Test in Adults

Getting an autism diagnosis as an adult has long meant navigating a patchwork of waiting lists, travel to specialist clinics, and assessment tools designed decades ago for children. Now, a preregistered retrospective study published in the Journal of Autism and Developmental Disorders offers some of the strongest evidence yet that a structured, telehealth-delivered autism assessment for adults can produce diagnostic decisions that two independent experts would broadly agree on. The finding matters because millions of adults—many of them women, people of color, or people who learned to camouflage their autistic traits in childhood—remain undiagnosed, and the pathways available to them are often unstandardized, expensive, or geographically out of reach.

The study, led by Eileen Crehan of the University of Massachusetts Chan Medical School together with Kelly Whaling, Hannah Jones, Jacqueline Shinall, and Rachel Loftin of the telehealth provider Prosper Health, set out to answer a deceptively simple question: if a blinded expert reviewer examined the same assessment materials that a treating clinician used, would the two reach the same conclusion about whether the patient met DSM-5-TR criteria for autism spectrum disorder? This kind of diagnostic agreement study does not by itself prove that a test is accurate, but it is a foundational step. If two professionals looking at the same information routinely disagree, no amount of accessibility or convenience can rescue the assessment from unreliability.

The sample consisted of 100 adults who primarily communicated through spoken language, drawn from routine clinical care at the telehealth organization. Critically, the researchers deliberately stratified recruitment so that at least half of the participants were Black, Indigenous, and people of color; the final sample achieved 51 percent BIPOC representation. That design choice is notable because autism research has historically over-relied on white, male, and childhood samples, and documented racial and ethnic disparities in autism identification—in children as well as adults—raise legitimate concerns about whether diagnostic instruments perform equitably across demographic groups. By building diversity into the sampling plan from the start, the team aimed to generate agreement estimates that were at least plausible to generalize beyond a narrow population.

The assessment protocol itself was unusual in both structure and philosophy. Evaluations generally included two 90-minute structured interviews spanning ten clinical domains mapped directly onto DSM-5-TR criteria, along with behavioral observation, standardized questionnaires, and collateral information from informants when available. The interview and the broader assessment pathway were developed with input from autistic adults, who gave feedback on interview content, question framing, sensory considerations, and the overall experience of being assessed. The team describes the approach as neurodiversity-affirming: rather than framing autism as a collection of deficits to be uncovered, clinicians were trained to affirm clients’ lived experiences while still applying explicit, criterion-level decision rules to determine whether diagnostic thresholds were met.

That combination—affirming stance plus rigid structure—is the study’s most provocative claim. Skeptics of neurodiversity-affirming practice have sometimes worried that reducing deficit-based language would soften diagnostic rigor, leading clinicians to rubber-stamp autism diagnoses. The authors push back directly on this idea, arguing that affirming clients’ experiences and setting aside deficit-based assumptions does not require presuming an autism diagnosis. The structured interview’s criterion-level decision rules functioned as guardrails: every clinical judgment had to be anchored to specific DSM-5-TR criteria, documented on structured rating forms, and traceable through the written record that the blinded reviewer would later examine.

So what did the numbers show? When the blinded expert—Crehan, who had no employment relationship with the telehealth organization and did not supervise any of its clinicians—reviewed the de-identified assessment materials, overall agreement with the original clinicians was 86.0 percent, with a 95 percent confidence interval running from 77.9 to 91.5 percent. Translated into Cohen’s kappa, the standard statistic for inter-rater agreement that corrects for chance agreement, the value was 0.684 with a confidence interval of 0.531 to 0.836. The authors had preregistered a kappa benchmark of 0.66 before accessing any data, and the point estimate cleared that bar—although, as the researchers candidly note, the confidence interval extends below it, meaning the data cannot rule out more modest agreement.

The fine-grained breakdown adds texture to the headline figure. Original clinicians classified 68 percent of the sample as autistic, while the blinded expert reviewer classified 66 percent—suggesting neither party showed a systematic inflation bias. Positive percent agreement, meaning the proportion of clinician-assigned autism diagnoses the expert confirmed, was 90.9 percent. Negative percent agreement, the proportion of non-autism classifications the expert confirmed, was 76.5 percent. In other words, when a clinician concluded that a patient was autistic, the independent reviewer very often agreed; disagreement clustered disproportionately in the smaller group of cases where clinicians concluded that criteria were not met. In a diagnostic context where a missed or delayed identification carries real consequences—lost access to accommodations, mental health supports, and self-understanding—that asymmetry is arguably the conservative kind of error to have.

Exploratory subgroup analyses hinted at lower agreement among BIPOC participants, although the authors are careful to stress that these estimates were imprecise, reflecting the limits of subgroup sample sizes. The finding echoes a long literature documenting racial and ethnic disparities in autism identification, including earlier diagnosis gaps and differential clinical profiles at referral, and it flags measurement equity as an open question rather than a settled one. The full R analysis code is publicly available on the Open Science Framework, an unusual degree of transparency for clinical research that allows independent analysts to scrutinize exactly how every number was produced.

The researchers are equally explicit about what the study does not show. Because the blinded reviewer relied on substantially the same information as the treating clinicians—interview transcripts, rating forms, questionnaires, and collateral reports—the design tests agreement under shared-information conditions, not diagnostic accuracy in the strict sense. A reviewer working from the same records as the original assessor could, in principle, inherit the same errors embedded in those records. Establishing criterion validity would require independent prospective validation: separate evaluations conducted without information overlap, multiple blinded reviewers, and ideally comparison against a reference standard. The authors call for exactly that next phase, and they also note the tension inherent in their own positionality: several authors are employees of the telehealth organization whose protocol was evaluated, and most of the authorship team identifies as neurodivergent. The study’s mitigation strategies—preregistration, random case selection, de-identification, a blinded non-employee reviewer, and transparent reporting of all deviations from the preregistered plan—were designed to counter both commercial and proximity-to-subject biases, but the dual roles are disclosed for readers to weigh.

For adults seeking assessment, the practical implications are nonetheless striking. Telehealth autism diagnosis accelerated during the COVID-19 pandemic and has since accumulated a growing evidence base in children, but adult-specific validation has lagged. Adults present particular challenges: decades of accumulated coping strategies, camouflaging behaviors that can mask core traits in structured observation settings, and co-occurring psychiatric conditions that complicate differential diagnosis. A structured, criterion-mapped interview delivered over video, developed with autistic adults and now shown to yield substantial agreement between independent raters, could meaningfully widen access to diagnosis for people for whom in-person specialist assessment is logistically or financially impossible. At the same time, the equity signals in the subgroup data and the shared-information limitation are reminders that accessibility and validity are separate questions, and that the field’s next task is to test these protocols against fully independent evaluations. If that work confirms the present findings, the default pathway to an adult autism diagnosis—long defined by scarcity, geography, and outdated instruments—may be on the verge of a genuine structural shift.

Subject of Research: Diagnostic agreement of a structured telehealth-based adult autism assessment mapped to DSM-5-TR criteria

Article Title: Diagnostic Agreement of a Structured Telehealth Adult Autism Assessment

Article References: Crehan, E., Whaling, K., Jones, H., Shinall, J., & Loftin, R. (2026). Diagnostic Agreement of a Structured Telehealth Adult Autism Assessment. Journal of Autism and Developmental Disorders. https://doi.org/10.1007/s10803-026-07571-2

Image Credits: AI Generated

DOI: 10.1007/s10803-026-07571-2

Keywords: autism, telehealth, adult diagnosis, DSM-5-TR, diagnostic agreement, Cohen's kappa, neurodiversity-affirming care, BIPOC health equity, psychological assessment, preregistration, Diagnostic, Agreement