A moving picture of brain activity
A systematic review has linked autism spectrum disorder (ASD) with more rigid and imbalanced patterns of brain-network activity during rest. Across seven eligible studies, researchers used Hidden Markov Models, or HMMs, to follow how resting-state fMRI signals moved among recurring brain states. That made it possible to compare how long states lasted, how much of the scan they occupied and how often the brain shifted between them.
The pooled results suggested that these HMM-derived measures could distinguish the groups represented in the studies. The pooled log odds ratio was 2.86, and the area under the curve (AUC), a summary measure of classification performance, was 0.85. The 95% confidence interval for the AUC ran from 0.72 to 0.98. The authors describe HMMs as promising for personalized diagnostics, not as a finished clinical tool.
Variation across the underlying studies was substantial. The heterogeneity statistic was 92% for the log odds ratio and 96.9% for the AUC. Heterogeneity measures how much results differ across studies, so the pooled numbers should be read as a summary of uneven evidence rather than as one performance level guaranteed to hold in a new setting.
The review team searched PubMed, Scopus and Web of Science in May 2025. Screening followed the PRISMA 2020 guidance, with independent review and consensus resolution of disagreements. To qualify, a paper had to be peer-reviewed and in English, apply HMMs to resting-state fMRI in people with ASD, and report classification performance or state metrics. Seven studies met the criteria.
The brain states did not last as long or as often
One of the main pooled state findings concerned the default mode network, or DMN, which the review places among the brain's integrative networks. Compared with typically developing controls, people with ASD had shorter mean lifetimes in DMN-associated states and longer mean lifetimes in sensory or attention-related hyperactivation states. Mean lifetime refers to the average duration of a state episode. The pooled Hedges' g was -4.19 for DMN-associated states and 3.80 for sensory/attention states. The latter outcome was rated high certainty.
Fractional occupancy, the share of the scan accounted for by a particular state, was also substantially lower in DMN states in ASD. The pooled Hedges' g was -6.22, while the I-squared heterogeneity statistic reached 99.6%. The review rated this outcome low certainty.
The narrative findings pointed in the same broad direction. They described reduced fractional occupancy and shorter mean lifetimes in DMN-hypersynchrony states, together with greater occupancy in sensory-motor and attention states. That pattern was replicated across two studies, even though the studies differed in their atlas choices, the number of HMM states and their participant samples.
The order in which states appeared also differed. Narrative analyses indicated fewer transitions from sensory-related states to DMN-related states and more self-transitions in sensory-motor states. A self-transition means the model remains in the same state rather than moving to another one. The review did not combine these transition probabilities into a single pooled estimate because the studies used incompatible state taxonomies.
A research signal, not a clinical test
The review also found a link between HMM-derived metrics and ADOS scores. The pooled correlation was -0.20, with a 95% confidence interval from -0.27 to -0.13. The result was rated high certainty, and the reported I-squared heterogeneity was 0%. A negative correlation means the two measures tended to move in opposite directions in the analyzed data; it does not show that altered brain states cause clinical differences.
Taken together, the findings describe an association between ASD and a different balance and timing of resting brain states. The authors interpret that pattern as reduced engagement of integrative networks and dominance of sensory states, and say HMMs could provide mechanistic insight and support personalized diagnostics. The analysis does not show that HMM use itself improves diagnosis, and it does not resolve the variation between studies.
Overall risk of bias across the evidence was low to moderate. Incomplete adjustment for confounders was the most common limitation. High heterogeneity affected both diagnostic accuracy outcomes and several state metrics, while the DMN fractional-occupancy result carried low certainty. These caveats are why the authors frame HMMs as promising for personalized diagnostics despite heterogeneity, rather than as a finished clinical approach.
The article was published as a peer-reviewed version of record on 21 August 2026. The review was registered in PROSPERO as CRD420251057196. The authors report no organizational support for the submitted work and declare no relevant competing financial or non-financial interests.
Paper data and sources
Original title: Autism Spectrum Disorder (ASD) Through the Lens of Hidden Markov Models (HMMs) Applied to Resting-State fMRI (rs-fMRI): A Systematic Review and Meta-analysis.
Authors: Mohamed Eltalkhawy, Amro K Barakat, Omar Ahmed Abdelaal et al.
Journal/Repository: Journal of autism and developmental disorders
Status: Peer-reviewed
First online: 2026-08-21
DOI: 10.1007/s10803-026-07358-5
Original paper · Full text