The Advanced Research Projects Agency for Health, known as ARPA-H, launched a new research program called SPECTRA on September 17, 2026, aiming to overhaul how autism spectrum disorder is diagnosed, understood, and treated using AI-driven analysis of biological, clinical, and behavioral data. The name stands for Systems for Phenotypic Evaluation, Clinical Trajectories, Response, and Agency, and the agency describes it as an effort to move autism care away from checklist-based diagnosis and toward individualized developmental trajectories.
Why Diagnosis Has Stayed Stubbornly Slow
Autism is currently diagnosed through structured behavioral observation, questionnaires filled out by parents and clinicians, and standardized assessments that can take months to schedule and complete, particularly in areas with few developmental pediatricians. The average age of autism diagnosis in the United States still lags well behind the age at which reliable signs are often present, delaying access to early intervention services that research consistently shows work best when started young. SPECTRA is built on the premise that combining genetic, physiological, video, voice, and electronic health record data with machine learning could compress that diagnostic timeline dramatically.
What the Program Will Actually Fund
ARPA-H’s announcement describes SPECTRA as bringing together biological, clinical, behavioral, and real-world data with advanced computational tools to enable earlier and more accurate diagnosis, identify the factors shaping individual developmental trajectories, and build more precise, individualized approaches to care. ARPA-H has organized the effort into four technical areas: one focused on identifying factors behind different developmental and health trajectories, a second on building privacy-preserving, federated AI models that can analyze data across institutions without centralizing it, a third on individualized selection and testing of interventions, and a fourth on technologies meant to improve communication, quality of life, and measurement of real-world function. Proposals for the fourth technical area are due October 18, 2026, with submissions opening October 4; the first two technical areas share a December 2, 2026 deadline with submissions opening November 18; and the third technical area’s proposals are due January 18, 2027, opening January 4. That mirrors a wave of academic research already underway: multiple 2026 studies have tested AI models analyzing facial expressions, eye-tracking patterns, vocal characteristics, and multimodal combinations of all three, with some reporting accuracy rates above 90% on curated research datasets.
The Psychologist Leading the Effort
ARPA-H tapped experimental psychologist Russell Shilling, a Navy veteran who previously worked at the Defense Advanced Research Projects Agency, to lead SPECTRA. Shilling, whose two sons are both on the autism spectrum, has framed the initiative around a lifetime-of-care approach that treats diagnosis as one point along a much longer arc rather than a single finish line. Earlier in his career he spent six years developing content for Sesame Street, built virtual reality exposure therapy used to treat Iraq War veterans with PTSD, and helped create some of the earliest chatbot-based tools designed to destigmatize seeking mental health treatment in clinical settings. Colleagues say that mix of military research experience, children’s media work, and personal familiarity with autism shapes how he is steering a program meant to bridge advanced computation and everyday family life.
A Field That Has Been Burned by Overconfident Claims Before
Researchers who study AI diagnostic tools are quick to note the gap between lab performance and real-world reliability. A 2026 scoping review of AI-based autism screening methods found that high accuracy figures often come from small or curated datasets, internally derived labels, or cross-validation performed within a single dataset, meaning a tool that looks excellent in a research paper can perform far worse on a diverse population it has never seen. Advocacy groups within the autistic community have also pushed back against framing autism purely as something to be detected and minimized, arguing that AI tools should be built to support individualized care and accommodation rather than solely to flag deviation from a neurotypical baseline.
Government Investment Signals a Shift in Priorities
ARPA-H, modeled loosely on the Pentagon’s DARPA and created to fund high-risk, high-reward biomedical research, has increasingly directed money toward AI-driven diagnostics across multiple conditions. SPECTRA’s launch suggests autism has become a priority category, likely reflecting both the rising prevalence of diagnoses, now affecting an estimated one in 31 children according to the CDC’s most recent surveillance data, and the enormous downstream cost of delayed intervention to families and school systems.
What to Watch For Next
The program is structured as a multi-year research initiative rather than a product launch, so parents should not expect an AI autism test at their pediatrician’s office anytime soon. The more immediate signal to watch will be which universities and companies apply for and win SPECTRA funding once the technical-area submission windows open this fall and winter, since that will indicate whether the program prioritizes screening speed, treatment personalization, or both. Longer term, if SPECTRA’s federated, privacy-preserving approach to combining data across institutions proves out, it could become a template ARPA-H applies to other developmentally complex conditions where diagnosis currently depends heavily on subjective clinical judgment rather than measurable biomarkers. Autism researchers who have watched previous federally funded diagnostic initiatives stall at the pilot stage say the real test for SPECTRA will be whether ARPA-H can get competing academic labs and private AI companies to share data and validation results openly, rather than each building a proprietary model in isolation, since fragmented, non-comparable datasets have historically been one of the biggest obstacles to producing an autism screening tool that works reliably outside the institution that built it.
Photo: Kampus Production / PEXELS via Pexels
