Chapter IX EHR-Defined Subtypes of Autism in Children and their Associations with Structural MRI
3. Material and Methods
4.2. Brain volume models
was a primary code for C3 and a secondary code for C2; C3 included several more epilepsy-related secondary codes. Unique from other clusters, C3 had both primary and secondary codes for speech and physical therapies, and several secondary codes related to digestive disorders.
Table IX-2 Primary conditions and procedures associated with each EHR cluster (C)
C Prevalence ProCode Category Prevalence PheCode Category
0
0.682 audiometry Ophthalmologic and otologic
diagnosis and treatment 0.682 Speech and language disorder mental disorders 0.682 Psychiatric
evaluation
Psychological and psychiatric evaluation and therapy
0.591 Developmental delays and disorders
mental disorders 0.636 otologic test Ophthalmologic and otologic
diagnosis and treatment 1
0.657 genetic testing Pathology 0.686 Delayed milestones endocrine/
metabolic 0.600 venipuncture Other therapeutic procedures 0.543 Speech and language disorder mental
disorders
2
0.703 EEG Electroencephalogram (EEG) 0.568 Convulsions neurological
0.595 hematologic tests Laboratory - Chemistry and
Hematology 0.514 Other headache syndromes neurological
0.568 compound
specific blood test Laboratory - Chemistry and Hematology
0.514 basic blood tests Laboratory - Chemistry and Hematology
3
0.967 otologic test Ophthalmologic and otologic
diagnosis and treatment 0.900 Speech and language disorder mental disorders 0.933 audiometry Ophthalmologic and otologic
diagnosis and treatment 0.767 Developmental delays and
disorders mental
disorders 0.933 acoustic test Ophthalmologic and otologic
diagnosis and treatment 0.733 Delayed milestones endocrine/
metabolic 0.867 compound
specific blood test Laboratory - Chemistry and
Hematology 0.600 Convulsions neurological
0.867 venipuncture Other therapeutic procedures 0.567 Lack of normal physiological
development, unspecified endocrine/
metabolic 0.767 EEG Electroencephalogram (EEG) 0.533 Epilepsy, recurrent seizures,
convulsions neurological
0.767 hematologic tests Laboratory - Chemistry and Hematology
0.700 basic blood tests Laboratory - Chemistry and Hematology
0.667 Psychiatric evaluation
Psychological and psychiatric evaluation and therapy
0.533 Speech Therapy Other diagnostic procedures (interview, evaluation, consultation) 0.533 genetic testing Pathology
0.500 speech treatment Other physical therapy and rehabilitation
associations between region volume and the EHR clusters; none of these associations survived multiple comparisons correction. Five of these six regions were in the left hemisphere. Three of the regions were in the basal ganglia (Figure IX-4): left accumbens area, left pallidum, and left putamen. The remaining three regions included the cerebellar vermal lobules (CVL) VI-VII, left temporal pole, and left transverse temporal gyrus (TTG) (Figure IX-5).
The left accumbens exhibited a diverging slope based on sex, with male region volumes increasing with age and female volumes decreasing; all other regions show similar trends of increasing or no change in volume with age, regardless of sex. Regions in the basal ganglia showed more dramatic associations with age compared to the other three regions. For all regions except the left temporal pole, C3 had the lowest intercept and C1 had the highest intercept. In the left temporal pole, C1 and C3 had nearly equivalent intercepts, lower than both C0 and C2.
Figure IX-4 General linear models of basal ganglia region volumes as a function of age, sex, and EHR-derived clusters. All three regions shown had weakly significant (p<0.05) associations with EHR clusters. Models are shown split by sex (right: Male, left: Female) for clarity.
Figure IX-5 General linear models of temporal lobe and cerebellum region volumes as a function of age, sex, and EHR-derived clusters. All three regions shown had weakly significant (p<0.05) associations with EHR clusters. Models are shown split by sex (right: Male, left: Female) for clarity.
5. Discussion
In this work, we identified four unique EHR sub-types in autistic children aged 2-10 and examined differences in brain region volumes across the identified sub-types. Three of these sub-types (C0, C1, C2) had well-defined EHR signatures, while the fourth (C3) had a more complex EHR structure. C0 contained patients with cognitive delays and auditory dysfunction; C1 contained patients primarily with physiological development delays; C2 contained patients with convulsions and perhaps mild epilepsy; and C3 contained patients with both epilepsy and a constellation of other conditions that overlapped with the other three clusters in specific ways.
Though statistically weak, the brain region associations found in our analysis were still intriguing given the characteristics of the identified EHR clusters. One distinguishing factor across clusters was the prevalence of convulsions and epilepsy, with clusters varying from no prevalence (C1) to high prevalence (C3). The accumbens, pallidum, putamen, and CBV VI-VII are all regions involved in movement regulation
[316], [317]; for all four regions, C1 had the highest region volumes while C3 had the lowest (Figure IX-4, Figure IX-5). This suggests that decreased volume in these regions may contribute to the presentation of convulsions/epilepsy experienced by autism patients. Similarly, another distinguishing characteristic across clusters was the prevalence of auditory dysfunction, which was elevated for clusters C0 and C3. The left TTG is believed to be part of the early auditory processing pathway [318]; Figure IX-5 showed that, again, C3 tended to have lower volumes in this region compared to all other clusters, but C0 had higher volumes than both C3 and C2. This suggests that C0 and C3 may represent varied subtypes of hearing loss, with patients in C3 linking to decreased volume in the left TTG and patients in C0 linking to other auditory system dysfunction outside of the brain. Other EHR patterns potentially support these diverging causes, with higher prevalence of speech and language therapies seen in C3 compared to interventions such as tympanastomy in C0.
Despite these points of interest, none of the identified brain regions passed multiple comparisons correction, indicating that these may have been spurious associations occurring by chance. This potential lack of brain associations may be surprising given the cognitive focus of the identified autism sub-types and the high incidence of significant associations between brain volume and autism in the literature [303], [306], [308], [309]. Perhaps these weak findings should not be surprising, however, given the small size of many autism studies which report region-specific differences [304], [319]. Larger studies of autism brain morphology tend to find fewer brain abnormalities [320]. Alternatively, it could be that region volume was not an appropriate measure for this application; many other anatomical features affect cognitive function, including gray matter thickness, gyrification, and white matter connectivity. Other studies have begun to investigate the potential associations between these alternative measures of brain structure and autism [321]–[323].
This article presented a novel analysis framework for combining information from EHR and MRI in a clinical setting. As such, it could be used in other cognitive disorders to uncover sub-types of EHR progression and corresponding brain phenotypes. The potential for this framework is not, however, limitless. Acquiring sufficiently sized joint EHR and MRI datasets is challenging and expensive. Such datasets are naturally biased towards “sicker” populations; to be included, patients must have sufficient EHR data and a clinical T1w MRI, both of which indicate acute or chronic medical problems. In addition, EHR data is typically sourced from a single medical system; any events occurring outside of that medical system are unknowable, potentially leading to noisy and unreliable EHR clusters. Many disorders, including autism, are also influenced by socio-economic and environmental factors; though the current framework does not account for such factors, it could potentially be extended to include more demographic information in the EHR clustering stage.