The first 5 pages, shown as text so search engines and readers can find this document.
Page 1
Preliminary testing of a
prespecified liability architecture
for autism: theory-guided
pathogenetic triad models
outperform strength-
matched alternatives
Darko Sarovic 1,2,3,4,5*
1Institute of Clinical Sciences, University of Gothenburg, Gothenburg, Sweden, 2Department of
Radiology, Harvard Medical School, Massachusetts General Hospital, Boston, MA, United States,
3Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts
General Hospital, Charlestown, MA, United States, 4Paediatric Health Professions and Paediatric
Radiology, Sahlgrenska University Hospital, Gothenburg, Sweden, 5Department of Women’s and
Children’s Health, Uppsala University, Uppsala, Sweden
Background: Neuropsychiatric conditions are heterogeneous and mechanistically
diverse, yet predictive modeling studies commonly aggregate features without
evaluating prespecified liability architectures. The Pathogenetic Triad (PT) is a
multilevel framework proposing that diagnostic outcomes reflect the joint
configuration of a trait-related domain, cognitive capacity (CC), and
neuropathological burden (NB). In autism, the trait-related domain corresponds
to autistic personality (AP). We evaluated this framework using out-of-sample
prediction as a structured empirical test of architectural coherence rather than as
a purely data-driven exercise in classification.
Methods: We analyzed a case–comparison cohort (n = 42, 21 autistic) with dense
multimodal characterization, including behavioral phenotyping, psychometric
measures, autonomic physiology, structural MRI morphometry, and
magnetoencephalographic indices. AP was indexed by the Autism-Spectrum
Quotient, CC by Wechsler’s scales of intelligence, and NB by heart-rate
variability as a proxy indicator. This multimodal structure enabled direct
comparison of theory-constrained PT models with strength-matched, domain-
restricted atheoretical combinations. Predictive discrimination was evaluated
using leakage-free nested cross-validation with permutation inference.
Results: Across multiverse specifications, low-dimensional PT models ranked
among the strongest models of comparable size and achieved discrimination
broadly comparable to higher-dimensional models within this dataset. Matched
comparisons indicated that including all three PT domains was associated with
systematic advantages relative to alternatives with similar univariate input
strength, consistent with complementary configurational information relating to
case status.
Conclusions: These findings provide preliminary evidence supporting the
Pathogenetic Triad as a multilevel architecture of autism liability. Although
Frontiers in Psychiatry frontiersin.org01
OPEN ACCESS
EDITED BY
Antonio M. Persico,
University of Modena and Reggio Emilia,
Italy
REVIEWED BY
Wenda Liu,
The George Washington University,
United States
Helena Brentani,
University of São Paulo, Brazil
*CORRESPONDENCE
Darko Sarovic
darko.sarovic@gu.se
RECEIVED 24 March 2026
REVISED 29 April 2026
ACCEPTED 19 May 2026
PUBLISHED 06 July 2026
CITATION
Sarovic D (2026) Preliminary testing of a
prespecified liability architecture for
autism: theory-guided pathogenetic
triad models outperform strength-
matched alternatives.
Front. Psychiatry 17:1837909.
doi: 10.3389/fpsyt.2026.1837909
COPYRIGHT
© 2026 Sarovic. This is an open-access
article distributed under the terms of the
Creative Commons Attribution License
(CC BY). The use, distribution or
reproduction in other forums is
permitted, provided the original
author(s) and the copyright owner(s) are
credited and that the original publication
in this journal is cited, in accordance
with accepted academic practice. No
use, distribution or reproduction is
permitted which does not comply with
these terms.
TYPE Original Research
PUBLISHED 06 July 2026
DOI 10.3389/fpsyt.2026.1837909
Page 2
based on a small and demographically restricted cohort, this densely
characterized dataset permitted explicit comparison of theory-guided and
atheoretical model spaces. The results illustrate how prespecified multilevel
frameworks can be operationalized and empirically evaluated in
neuropsychiatric samples.
KEYWORDS
autism, autistic traits, heart rate variability (HRV), liability architecture, multilevel
framework, multiverse analyses, predictive modeling, pathogenetic triad
Introduction
Autism spectrum disorder is a neurodevelopmental condition
characterized by differences in social communication, restricted and
repetitive behaviors, and atypical sensory processing (3). However,
autistic presentations are highly heterogeneous, varying in symp-
tom profile, cognitive ability, adaptive functioning, and co-occur-
ring psychiatric and medical conditions (4, 5). This heterogeneity
complicates formulation of mechanistic models that capture com-
monality while respecting phenotypic diversity.
In response to this heterogeneity, biomarker discovery and
multivariable classification efforts have accelerated (6, 7). Yet
many models are assembled in a largely data-driven way without
an explicit mechanistic framework constraining feature selection or
structure. This can yield high-dimensional models that are difficult
to interpret and generalize, and hard to translate into testable
hypotheses (8–10). A theory-guided feature space can impose
structure, improve parsimony and interpretability, and link predic-
tive performance to liability architecture, potentially supporting
clinical uptake.
These limitations reflect two gaps. First, most mechanistic
accounts of autism focus on a single level of analysis and rarely
specify how levels should be linked, limiting their value as global
explanations. Second, even when multilevel theories are proposed,
they are rarely operationalized into specific variables and quantita-
tive predictions suitable for multivariable models (11, 12). As a
result, the field lacks a widely adopted multilevel framework that is
directly testable in predictive evaluation pipelines.
A related issue concerns which features enter predictive models.
Markers that are relatively homogenous across autistic individuals
may yield small but stable effects, whereas mechanisms prominent
in only subsets may produce larger effects but generalize poorly.
Without a principled account of autism’s latent structure, it is
unclear how to combine these signals in a way that is both
mechanistically meaningful and statistically robust. More generally,
even sophisticated algorithms can only operate within the feature
space they are given–so defining that space is itself a theoretical act.
We refer to this as theory-guided predictive testing: using a
multilevel framework to map observed variables onto a small set
of interpretable latent dimensions and then evaluating whether this
prespecified architecture carries non-redundant information
aligned with case-status. The goal is not to optimize a classifier
but to test whether the proposed liability structure survives empir-
ical testing.
The Pathogenetic Triad (PT) is a multilevel framework for
neurodevelopmental liability (13, 14) linking genetic, biological,
cognitive and behavioral measures through measurable compo-
nents and mechanisms. Although formulated as a neurodevelop-
mental liability framework, the present study evaluates the PT using
autism as a specific and well-characterized test case. The model
proposes that autistic outcomes reflect the configuration of three
interacting dimensions: Autistic Personality (AP), Cognitive
Capacity (CC), and Neuropathological Burden (NB). AP indexes
autistic traits; CC captures cognitive resources that can amplify or
contain the behavioral expression of AP; and NB reflects neurobi-
ological vulnerability that may constrain compensatory capacity.
Within the PT framework, NB refers to burden-related develop-
mental constraint rather than to any single downstream biomarker.
Individual biomarkers may reflect aspects or consequences of such
constraint, but the construct is not represented by any single
marker. In the present pilot test, NB was operationalized using
short-term vagally mediated heart-rate variability (HRV), as a
pragmatic proxy indicator rather than a direct measure of neuro-
pathology, because autonomic flexibility is theorized to modulate
how biological and psychosocial stressors are buffered and trans-
lated into developmental constraint, thereby limiting compensatory
capacity over time (15, 16). The framework yields testable postu-
lates, including that autistic and non-autistic individuals occupy
different regions in triad space and that joint triad-based models
should capture configurational information beyond isolated indi-
cators (assessed through the first two questions below).
To our knowledge, PT-guided models have not been tested
within formal predictive evaluation pipelines, and few multilevel
theories have been evaluated as explicit, operationalized feature
spaces for multivariable prediction. If AP, CC, and NB approximate
latent dimensions organizing autistic and non-autistic individuals,
then a small set of triad-based indicators should capture much of
the diagnostic signal; alternative operationalizations should per-
form similarly; and high-dimensional models should add only
Abbreviations: AP, Autistic Personality; CC, Cognitive Capacity; NB,
Neuropathological Burden; PT, Pathogenetic Triad; TI, TriadIndex; AQ,
Autism-Spectrum Quotient; ECG, electrocardiogram; FSIQ, Full-Scale IQ;
HRV, heart rate variability; IQ, intelligence quotient; SD1, Poincaré SD1
(short-term HRV metric); WAIS-IV, Wechsler Adult Intelligence Scale, Fourth
Edition; WMIQ, Working Memory Index (WAIS-IV); AUC, area under the
receiver operating characteristic curve; BA, balanced accuracy; CI, confidence
interval; CV, cross-validation; LOOCV, leave-one-out cross-validation; M, mean;
ROC, receiver operating characteristic; SD, standard deviation; ANAT, anatom-
ical domain; BEH, behavioral domain; MEG, magnetoencephalography domain.
Sarovic 10.3389/fpsyt.2026.1837909
Frontiers in Psychiatry frontiersin.org02
Page 3
modest incremental value beyond the triad. Observing this pattern
would support the PT as a plausible liability framework. We treat
predictive modeling as a test of mechanistic structure rather than
pattern recognition. Accordingly, we emphasize relative model
ranking, stability across multiverse specifications, and robustness
under strict leakage control, rather than interpreting performance
estimates as stable population parameters. We evaluate this logic in
a case–comparison cohort using four interrelated questions.
Does a prespecified PT operationalization show out-of-
sample separation? Specifically, does a simple, low-dimensional
PT representation capture discriminative structure aligned with
case status under conservative, leakage-free evaluation?
Do theory-constrained triad-models outperform data-driven
alternatives? If the framework approximates core liability dimen-
sions, PT-consistent models should rank among the highest-per-
forming models of similar size relative to atheoretical combinations
matched on univariate input strength.
How parsimonious is the PT representation of autism risk?
To what extent can a low-dimensional, theory-constrained repre-
sentation recover the discriminative signal observed in higher-
dimensional, atheoretical benchmarks?
Can structured mapping of PT-consistent operationalizations
provide a blueprint for preregistered cumulative testing
across samples?
Together, these analyses evaluate the PT as an empirically useful
multilevel framework and illustrate a practical approach to theory-
guided predictive testing.
Materials and methods
Participants and design
We analyzed an existing IQ-matched case–comparison cohort of
autistic and non-autistic men that has been described elsewhere (17,
18). Autistic participants were recruited from existing longitudinal
studies where they had received a diagnosis of autism spectrum
disorder on at least two separate occasions, using full diagnostic
neuropsychiatric multidisciplinary evaluations, including the Autism
Diagnostic Observation Schedule. Individuals with other neuropsy-
chiatric diagnoses, such as Attention-Deficit/Hyperactivity Disorder,
were excluded. Non-autistic participants were recruited from
Gothenburg, Sweden using flyers and on university campus, and
individuals with any self-reported psychiatric diagnosis were ex-
cluded. The groups were matched for full-scale IQ (FSIQ) and
participants with a FSIQ below 80 were excluded.
Analyses used complete-case subsets defined by the predictors
required for each model family. The full cohort (n = 42; 21 autistic) was
used for the PT-complete analyses (primary PT model and
TriadIndex). Analyses requiring the expanded predictor set used the
largest complete-case subset across those variables (n = 35; 17 autistic).
The study was approved by the regional ethical board in
Gothenburg, Sweden (DNR: 552-14) and all participants provided
written informed consent.
Measures
PT operationalization and study-specific
operationalization
The PT is a theoretical framework defined at the level of latent
domains. The present study does not claim to measure those
domains exhaustively. Rather, it tests a pre-specified low-dimen-
sional operationalization in which AQ, WMIQ, and HRV-derived
metrics serve as indicator-level proxies for AP, CC, and NB within
this dataset.
AP was indexed using the Autism-Spectrum Quotient total
score (AQ) (19). It was used here as a pragmatic, continuous trait-
level proxy for AP, but AP as formulated in the PT is broader than
any single questionnaire score. AQ total may underrepresent
aspects of AP, and some items may capture variance related to
cognition or internalizing symptoms rather than AP per se.
CC was assessed with the working memory subscale (WMIQ) of
Wechsler Adult Intelligence Scale (WAIS-IV) (20). WMIQ is a
plausible partial proxy for compensatory capacity, but compensation
is multi-faceted and unlikely to be exhausted by a single instrument.
Broader executive, language-mediated, adaptive, and social-cognitive
resources also contribute to compensation. Additional IQ indices
were evaluated in exploratory multiverse analyses.
NB was operationalized using cardiac HRV derived from ECG,
with lower vagal modulation interpreted as higher putative burden
(21). The primary NB measure was SD1, a Poincaré -plot short-term
vagally mediated HRV metric computed from interbeat intervals.
HRV is used here as a theory-guided proxy for burden-related
developmental constraint, not as a direct measure of neuropathol-
ogy. Resting vagal HRV is mechanistically plausible because auto-
nomic flexibility may index how biological and psychosocial insults
are buffered and translated into developmental constraint, but it is
state-sensitive and captures NB only partially. ECG preprocessing
and artifact handling are described in the Supplement. Additional
HRV metrics were evaluated in exploratory multiverse analyses.
For the primary 2D representation, CC and NB were combined
into a theory-guided composite vulnerability axis because, within
the PT framework, NB is hypothesized to constrain compensatory
capacity, and the two variables therefore jointly define a nonspecific
compensation-constraint dimension that modulates whether AP
translates into diagnosable impairment. This parsimonious 2D
projection facilitates visualization and low-dimensional testing,
but should not be interpreted as evidence that CC and NB form a
single empirically derived latent factor; accordingly, we also evalu-
ated the 3D PT model with CC and NB modeled separately.
Non-PT predictors
To position PT-based models within a broader, data-driven
“model space”, we included candidate predictors from three addi-
tional domains:
Behavioral (BEH): available self-report measures of alexithymia
(Toronto Alexithymia Scale-20) (22) and sensory processing
(Adult/Adolescent Sensory Profile total score and subscales) (23).
Sarovic 10.3389/fpsyt.2026.1837909
Frontiers in Psychiatry frontiersin.org03
Page 4
Anatomical (ANAT): selected structural magnetic resonance
imaging-derived (MRI) morphometry measures, including hippo-
campus and amygdala volumes and global grey- and white-
matter volumes.
Neurophysiological (MEG): discriminative magnetoen
cephalographic (MEG) markers of visual face processing (M130/
M170 amplitude and latency, and a late component amplitude).
Within the PT framework, anatomical and functional neurobi-
ological variables are considered developmental end-result bio-
markers and were therefore analyzed as separate neurobiological
benchmark domains rather than treated as interchangeable indica-
tors of NB. MRI and MEG preprocessing pipelines are presented in
(17, 18). Complete lists of included variables and their domain
assignments are provided in the Supplement.
Data processing and univariate
characterization
Continuous variables were screened for right-skewed non-
normality (Shapiro–Wilk p < 0.05; skewness > 0.5) and log-
transformed when indicated. Predictors were then oriented so
higher values reflected higher autism liability (if univariate area
under ROC (AUC) < 0.5), values were sign-flipped) and then z-
scored for descriptive summaries (for supervised models, z-scoring
was performed within each outer training set). Univariate strength
was summarized as AUC = max(AUC, 1–AUC), and model “input
strength” was defined as the mean AUC of included predictors.
Analytic overview
Analyses addressed the four prespecified questions described in
the Introduction. Except for the high-dimensional landscape,
models were evaluated using leakage-free nested leave-one-out
cross-validation (LOOCV) (24): each participant served once as
the held-out test case, and all preprocessing, feature construction,
and model fitting were confined to the corresponding outer training
set. Predictive evaluation used ridge-regularized logistic regression;
predictors were z-scored within each outer training set and applied
to the held-out participant. The ridge penalty (C) was tuned within
each outer training set via inner stratified cross-validation optimiz-
ing log loss. Performance was estimated from out-of-fold predic-
tions using AUC (discrimination), balanced accuracy (thresholded
performance), and the Brier score (calibration). Uncertainty was
quantified using bootstrap confidence intervals over paired out-of-
fold predictions and labels, and the primary model additionally
used label-permutation testing with the full pipeline refit per
permutation. Unless noted, comparisons within a model family
used complete-case subsets to avoid differences in missingness. No
prospective power calculation was performed because this conve-
nience pilot was designed for leakage-controlled out-of-sample
estimation rather than detection of small effects. Analyses were
performed in Python 3.11; dependencies and a reproducibility guide
(including scripts and synthetic data (1)) are provided in
the Supplement.
PT model performance
We evaluated a prespecified PT operationalization comprising
AQ as AP, WMIQ as CC, and SD1 as NB. Consistent with the PT
theory, we operationalized the primary two-dimensional PT deci-
sion space using standardized AP and an NB/CC composite axis
combining neuropathological liability and cognitive capacity, using
the mean of their z-scores (z(CC) + z(NB))/2.
Permutation and null performance. Statistical significance of
the model’s discrimination was assessed using label-permutation
testing, repeating the full nested CV pipeline across 20,000 permu-
tations to obtain a null distribution of AUC and an empirical
p-value.
Unsupervised structure check. As a descriptive check of struc-
ture in PT space, we applied K-means clustering (K = 2) to the full-
sample coordinates (z(AP), z(NB/CC)). Cluster labels were aligned
to diagnosis to maximize agreement, and the resulting operating
point (sensitivity/specificity) was reported descriptively.
The Triadindex as a continuous PT risk
score
We examined whether the triad’s information could be col-
lapsed onto a single continuous theory-guided risk axis. We
constructed a TriadIndex (TI) by computing the weighted sum of
the same three PT variables (standardized and risk-oriented) into a
single index (re-standardized across the raw TI) intended to reflect
overall PT-consistent liability. The primary TI used a theory-guided
weighting (emphasizing AP with a ratio of 2:1:1); an equal-weight
TI was evaluated as a theory-agnostic sensitivity variant. TI analyses
used complete cases with data for the PT variables (n = 42).
Constructing the triplet multiverse space
We constructed a multiverse of all triplet (k = 3) ridge logistic
regression models under a domain restriction: each triplet con-
tained three predictors drawn from three distinct domains (no
repeated domains within a triplet) from the six available domains
(AP, CC, NB, BEH, ANAT, MEG), preventing models from being
dominated by a single measurement domain. To ensure fair
comparisons, we included individuals with complete data across
the full predictor set (n = 35). We focused on k = 3 because triplets
are the model size that represents the full PT configuration while
remaining comparable to data-driven triplets of the same size.
Each model was annotated by its mean univariate AUC (input
strength), and by the number of PT-domains it contains: “PT
triads” were defined as triplets containing all three PT domains
(3PT); non-PT triads were defined by the number of PT domains as
0PT, 1PT and 2PT.
Matched paired comparisons of PT vs non-PT
triplets
To test whether PT triplets outperform data-driven alternatives
after controlling for univariate signal, we compared each 3PT triplet
Sarovic 10.3389/fpsyt.2026.1837909
Frontiers in Psychiatry frontiersin.org04
Page 5
to matched triplets from the 0PT pool and an AP-only 1PT pool.
For each 3PT triplet, comparator triplets were selected by nearest-
neighbor matching on model “input strength” (mean univariate
AUC); comparator performance was summarized as the mean AUC
across the K matched models (primary K = 5). Sensitivity analyses
used K = 1 and profile-based matching on the three predictor-level
AUC values with K = 5 (see Supplementary Methods). For each 3PT
triplet, we computed DAUC = AUC3PT − AUCmatched and tested for
a systematic advantage (DAUC > 0) across triplets using Wilcoxon
signed-rank tests. Analyses were run for 3PT vs 0PT and for 3PT vs
AP-only 1PT (to isolate added value beyond AQ) and summarized
in matched-effect plots.
Exploratory identification of candidate PT triplets
Within the same k = 3 multiverse results, we identified all PT
triads and ranked them by AUC. We also computed the median
AUC across all PT triads in which that predictor appeared to
indicate which one contained the most discriminative signal.
Comparing a parsimonious PT-model with
a high-dimensional benchmark
We contrasted parsimonious PT-based models (our primary
model as k = 2 and k = 3; 2D; 3D) against a high-dimensional ridge
logistic regression benchmark (k = 28; 28D) including the full
available predictor set in the expanded domain space.
High-dimensional landscape
To characterize how performance scales with model size in the
expanded predictor space, we generated a high-dimensional model
landscape by evaluating models of size k = 1 to k = 28 derived from
2,000 random orderings of the full multivariable predictor set. For
each random ordering, we formed nested models (first 1 predictor,
first 2 predictors, …, all predictors) and evaluated each model under
LOOCV using a fixed ridge penalty C selected from the tuned 28D
high-dimensional benchmark model to avoid embedding a full inner-
loop tuning within the landscape computation. This produces a
leakage-free description of the typical relationship between dimen-
sionality and cross-validated discrimination in the expanded predic-
tor space while keeping the landscape computationally tractable.
As an additional check focused on direct fairness for low
dimensional models, we also evaluated all k = 2–3 models (exhaus-
tive enumeration of all domain-restricted predictor combinations; n
= 318 and 1856 respectively), enabling direct comparison between
the PT models and theory-agnostic low-dimensional subsets from
the expanded predictor space.
Generative AI assistance
A large language model (ChatGPT 5.2: OpenAI) was used to
assist with language editing, clarity of exposition, code refactoring,
and identifying formatting errors. It was not used to generate
results, perform statistical analyses or interpret findings. No
model outputs were accepted without author verification.
Results
Triad variables and univariate diagnostic
resolution
The diagnostic resolution of all candidate predictors and the
group distributions for the three core PT components in the
primary model are presented in Supplementary Figure 1 and
Supplementary Table 2. Across the full predictor set, univariate
in-sample AUC ranged from 0.526 to 0.882.
For the prespecified primary model, there were group differences
between all three predictors (AQ Total: autism = 24.6 ± 8.9, non-autism
11.9 ± 6.0, p = 3×10-6, whole sample = 18.2 ± 9.7; WMIQ: autism =
103.3 ± 13.9, non-autism 114.0 ± 12.4, p = 0.01, whole sample = 108.6 ±
13.9; SD1: autism = 124.3 ± 63.1, non-autism 185.1 ± 53.9, p = 0.002,
whole sample = 154.7 ± 64.8). As expected given FSIQ matching, the
groups did not differ meaningfully in FSIQ (p = 0.27). Because AQ was
the strongest single predictor, we also evaluated an AQ-only compa-
rator model that, under the same leakage-free nested pipeline, yielded
AUC = 0.862, Brier = 0.163, and BA = 0.833 (Supplementary Table 3).
Does a prespecified PT operationalization
show out-of-sample separation?
Primary PT model in triad space
In two-predictor triad space, autistic and non-autistic partici-
pants occupied largely distinct regions, with an expected diagonal
decision boundary (Figure 1).
Discrimination was high: AUC = 0.934 (95% CI 0.845–0.995).
The BA-maximizing threshold yielded BA = 0.881 (95% CI 0.813–
0.978), with sensitivity = 0.952 and specificity = 0.810. The Brier
score was 0.114 (95% CI 0.069-0.165).
Permutation test and unsupervised clustering
The label-permutation test confirmed that observed perfor-
mance is highly unlikely under the null. The observed AUC lay
far above the 95th percentile of the null distribution (empirical
p = 5×10-5) (Supplementary Figure 2).
As a complementary unsupervised check, k-means clustering
showed above-chance alignment with case status (BA = 0.762, sen-
sitivity = 0.524, specificity = 1.000). The clustering operating point lay
on the ROC curve of the supervised PT model (Supplementary
Figure 2), consistent with structure in the triad space aligned with
case status in this cohort, even without access to labels.
TriadIndex: collapsing triad space onto a single
axis
The TI model achieved AUC = 0.943 (95% CI 0.857–1.000) and
BA = 0.893 (95% CI 0.767–1.000) with Brier = 0.108 (95% CI 0.063–
0.163). An equal-weight TI sensitivity model performed similarly
(AUC = 0.948 [0.872–0.995]). The results (triad index with decision
boundary, ROC curve and calibration curve) are presented in
Supplementary Figures 3, 4.
Sarovic 10.3389/fpsyt.2026.1837909
Frontiers in Psychiatry frontiersin.org05