Comorbidity Clusters in Autism Spectrum Disorders: An ...was characterized by multisystem disorders including gastrointestinal disorders (prevalence 24.3%) and auditory disorders and

Comorbidity Clusters in Autism Spectrum Disorders:An Electronic Health Record Time-Series Analysis

WHAT’S KNOWN ON THIS SUBJECT: Individuals with autismspectrum disorders have a higher comorbidity burden than thegeneral pediatric population, including higher rates of seizures,psychiatric illness, and gastrointestinal disorders.

WHAT THIS STUDY ADDS: Comorbidities do not occur evenly. Ourclustering analysis reveals subgroups characterized by seizure,psychiatric disorders, and complex multisystem disordersincluding auditory and gastrointestinal disorders. Correlationsbetween seizure, psychiatric disorders, and gastrointestinaldisorders are validated on a sample from a second hospital.

abstractOBJECTIVE: The distinct trajectories of patients with autism spectrumdisorders (ASDs) have not been extensively studied, particularly re-garding clinical manifestations beyond the neurobehavioral criteriafrom the Diagnostic and Statistical Manual of Mental Disorders. Theobjective of this study was to investigate the patterns of co-occurrence of medical comorbidities in ASDs.

METHODS: International Classification of Diseases, Ninth Revision codesfrom patients aged at least 15 years and a diagnosis of ASD wereobtained from electronic medical records. These codes were aggregatedby using phenotype-wide association studies categories and processedinto 1350-dimensional vectors describing the counts of the most commoncategories in 6-month blocks between the ages of 0 to 15. Hierarchicalclustering was used to identify subgroups with distinct courses.

RESULTS: Four subgroups were identified. The first was characterized byseizures (n = 120, subgroup prevalence 77.5%). The second (n = 197)was characterized by multisystem disorders including gastrointestinaldisorders (prevalence 24.3%) and auditory disorders and infections(prevalence 87.8%), and the third was characterized by psychiatric dis-orders (n = 212, prevalence 33.0%). The last group (n = 4316) could notbe further resolved. The prevalence of psychiatric disorders was un-correlated with seizure activity (P = .17), but a significant correlationexisted between gastrointestinal disorders and seizures (P , .001). Thecorrelation results were replicated by using a second sample of 496individuals from a different geographic region.

CONCLUSIONS: Three distinct patterns of medical trajectories were iden-tified by unsupervised clustering of electronic health record diagnoses.Thesemay point to distinct etiologieswith different genetic and environmen-tal contributions. Additional clinical and molecular characterizations will berequired to further delineate these subgroups. Pediatrics 2014;133:1–10

AUTHORS: Finale Doshi-Velez, PhD,a Yaorong Ge, PhD,b andIsaac Kohane, MD, PhDa

aCenter for Biomedical Informatics, Harvard Medical School,Boston, Massachusetts; and bCenter for Biomedical Informatics,Wake Forest University, Winston-Salem, North Carolina

KEY WORDSautism, seizure, psychiatric disorders, comorbidity, clustering

ABBREVIATIONSASD—autism spectrum disordersEHR—electronic health recordICD-9—International Classification of Diseases, Ninth RevisionPheWAS—phenotype-wide association studies

Dr Doshi-Velez designed and performed all of the analyses anddrafted and revised the manuscript; Dr Ge supplied data fromWake Forest University; Dr Kohane supplied data from theChildren’s Hospital Boston, provided guidance on theinterpretation of the analyses, and critically reviewed andrevised the manuscript; and all authors approved the finalmanuscript as submitted.

www.pediatrics.org/cgi/doi/10.1542/peds.2013-0819

doi:10.1542/peds.2013-0819

Accepted for publication Oct 22, 2013

Address correspondence to Finale Doshi-Velez, PhD, Center forBiomedical Informatics, Harvard Medical School, 10 Shattuck St,Boston, MA 02115. E-mail: [email protected]

PEDIATRICS (ISSN Numbers: Print, 0031-4005; Online, 1098-4275).

Copyright © 2014 by the American Academy of Pediatrics

FINANCIAL DISCLOSURE: Drs Doshi-Velez and Kohane werepartially funded through the Conte Center; and Dr Ge hasindicated he has no financial relationships relevant to thisarticle to disclose.

FUNDING: All phases of this study were supported by theInformatics for Integrating Biology and the Bedside NIH #2U54LM008748. Dr Doshi-Velez is supported by the National ScienceFoundation under a CI TraCS grant awarded in 2012.

POTENTIAL CONFLICT OF INTEREST: The authors have indicatedthey have no potential conflicts of interest to disclose.

PEDIATRICS Volume 133, Number 1, January 2014 1

ARTICLE

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

Clinical manifestations of autism spec-trum disorders (ASDs) beyond the coreDiagnostic and Statistical Manual ofMental Disorders criteria have beengaining increasing attention.1–5 With theprevalence of autism near 1%,6 un-derstanding this comorbidity burden isespecially important (not least becauseof the clinical resources this burdenentails). Understanding the co-occurrencepatterns among comorbidities in ASD isthe first step for uncovering the un-derlying etiologies associated with ASDand stratifying the risk of various con-ditions across individuals with ASD.

Previous research has quantified theprevalence of various comorbidities inASD, including the rates of gastroin-testinal disorders and complaints,7,8

epilepsy,9,10 sleep disorders,11 muscu-lar dystrophy,12–14 and psychiatric ill-ness.15 However, with the exception ofKohane et al,5 these studies have in-volved small samples (under 200 indi-viduals) and have focused on theprevalence of a single disorder. Thedistinct clinical trajectories of patientswith ASD have not been extensivelystudied, particularly as regards tothese comorbidities.

The objective of this study was to betterunderstand the co-occurrencepatternsof comorbidities in ASD. ASD is knownto be a heterogeneous disorder withcomplex genetic underpinnings.16

Subsets of these variants are commonto other diseases. For example, thereexistmutations common to ASD, attention-deficit/hyperactivity disorder, andschizophrenia.17–20 Mutations re-sponsible for Rett syndrome21 andFragile X22 carry a much higher riskof ASD and epilepsy. The associationbetween peripartum infections andASD is also well-documented.23–25 Sub-groups clustered on clinical criteriamay be enriched for different etiolo-gies; if so, these criteria may then beused to specifically target and evaluatetherapies or preventive measures.

Weuseelectronichealth records (EHRs)to cluster co-occurrence patterns ofclinical conditions. EHR and claims datahave large research potential to dis-cover both disease–disease and dis-ease–gene correletions.26–28 Studieshave described procedures for assess-ing patient similarities between pa-tients29,30; applications include riskstratification strategies for diabetesand cardiovascular disease.31 Carneyand Jones32 use claims data froma large provider to identify comorbid-ities associated with bipolar disorders.Particularly valuable is that these dataalready document the date of the clinicvisit during which a disorder wasreported.

Previouswork in clustering phenotypesin ASD has relied on surveys and di-agnostic tests, limiting the sample size.For example, Miles et al33 divide ASDinto 2 clusters, “essential” and “com-plex” based on the manifestation ofsignificant dysmorphology or micro-cephaly. They find that patients withcomplex ASDs have poorer outcomes,including lower IQ and more seizures.Other studies have focused on thecore neurobehavioral criteria. Wigginset al34 find clusters along disease se-verity, whereas Lane et al35 discoversensory processing subtypes. Otherstudies36–39 find clusters along cogni-tive, language, and behavioral criteria.Sacco et al40 find patterns among bothneurodevelopmental factors as well asimmune and circadian dysfunction.

Through the use of EHR data, we finddistinct clinical trajectories of comor-bidities outside of the core neuro-behavioral ASDDiagnostic andStatisticalManual of Mental Disorders criteria.From a sample of 4927 patients aged 15years from a tertiary-care pediatrichospital (mean follow-up 11 years, SD4.8 years), our clustering analysesrevealed 3 high-morbidity subgroups: 1characterized by seizures, 1 charac-terized by psychiatric disorders, and 1

characterized by more complex multi-system disorders. These phenotypicdistinctions may point to distinct eti-ologies with different genetic and en-vironmental contributions.

METHODS

Patients

We identified 13 740 individuals with atleast 1 International Classification ofDiseases, Ninth Revision (ICD-9) codefor ASD (299.00, 299.01, 299.80, 299.81,299.90, 299.91) by using infrastructurefrom the i2b2 National Center forBiomedical Computing41 at BostonChildren’s Hospital. Of these, 4934individuals (78% boys) were at least 15years old. Key patterns found amongthese individuals were examined for ina sample of 496 (80% boys) individualsfrom Wake Forest University HealthSciences (the full study could not bereplicated because of the small samplesize). The institutional review boardsof Harvard Medical School, BostonChildren’s Hospital, and Wake ForestUniversity reviewed and approved theresearch protocol.

Methods

The6905 ICD-9codespresent inourdatawere aggregated into 802 categoriesused in phenotype-wide associationstudies (PheWAS).42 Procedure codeswere ignored. The ICD-9 codes for keycategories are provided in the Sup-plemental Information. Individuals withmore than 50 instances of the samecategory code in a 6-month periodwere excluded because their recordswere dominated by conditions unrelatedto ASD (eg, renal failure, oncology visits).Finally, we only considered categoriesthat had at least 5% prevalence in thesample. This pre-processing resulted in45 common category codes and 4927individuals.

Foreachpatient,weconstructeda time-series with 30 6-month windows from

2 DOSHI-VELEZ et al

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

birth to age 15. For each time window,we counted the number of occurrencesof each of the 45 categories for thatpatient in that window. This processingstep resulted in a 30 3 45 = 1350-dimensional vector of counts per pa-tient. Hierarchical clustering by usinga Euclidean distance and Ward’smethod resulted in 4 clusters, whereclusters were constrained to a mini-mum size of 2% of the overall sample.

Differences between the clusters wereassessed by using permutation tests.We computed a x2 statistic comparingthe expected number of code counts ineach time window in each cluster tothe observed numbers for each pa-tient. Next, the same statistic wasrecomputed on 15 000 permutations ofthe patients. Thus, the permutationspreserved the cluster sizes and theintrapatient comorbidity statistics. Theempirical P value for each cluster wascomputed by comparing the x2 statis-tic for our clustering to the empiricaldistribution created from the permu-tations. The number of permutationswas chosen to estimate small P valuesto sufficient precision to then applya Bonferroni multiple hypothesis cor-rection to these P values.

Analyses were performed by usingMatlab 7.14.0.739 (Mathworks; Natick,MA) (R2012a) and R 2.15.2 (R Founda-tion; Vienna, Austria).

RESULTS

Thehierarchical clusteringresulted in4clusters (Table 1). Three clusters weresmall (n = 120; 197; 212), and the finalcluster (n = 4316) could not be furtherresolved into subgroups. Eighty-twooutliers were excluded because theywere far from all the other clusters andeach other. Individuals in the smallersubgroups averaged over 5 timesmorecodes than the larger subgroup, sug-gesting that the smaller subgroupshad higher morbidity than the overallsample. Subgroup 1 had slightly fewer

boys and the oldest age of first di-agnosis. Individuals in subgroups 1 and2 had more diagnoses of autism thanAsperger syndrome, whereas sub-group 3 had more individuals withAsperger syndrome.

Subgroup characteristics are summa-rized in Table 2 and Fig 1 (the ICD-9codes used to define these comorbid-ities are in the Supplemental Information).Subgroup 1 was characterized by sei-zures (prevalence 77.5%). Subgroup 2was characterized by multisystem dis-orders including gastrointestinal dis-orders (24.4%) and early ear infectionsand auditory disorders (87.8%). Sub-group 3 was characterized by psychi-atric disorders (prevalence 33.0%). Allof these subgroups had higher levels ofcardiac disorders (30.8%, 33.0%, and24.0%, compared with 6.9%) and in-tellectual disability (60.0%, 48.7%, and27.8%, compared with 12.7%). The finalsubgroup had no significantly elevatedcomorbidities, so we now focus on the3 high-morbidity subgroups.

The temporal patterns of develop-mental delays varied between thesubgroups. Individuals in subgroup 2,characterized by multisystem disorders,

had a spike in diagnoses for devel-opmental delays (ICD-9 codes startingwith 315) at age 2.5 (Fig 2). Specificdevelopmental delays for individuals insubgroup 1 rose steadily through age5. In contrast, subgroup 3 had steadyand relatively low prevalence of spe-cific developmental delays throughage 15. Table 3 shows the specific ICD-9 codes contributing to the devel-opmental delays for each subgroup.Codes for expressive language disor-der (315.31) were more common insubgroups 2 and 3, and codes formixed developmental disorder (315.5)were more common in subgroups 1and 2. These patterns are consistentwith higher proportion of autism insubgroups 1 and 2 and Asperger syn-drome in subgroup 3.

Finally, the rates of visits where ASDdiagnoseswere recorded varied widelybetween the different subgroups(Fig 3). Individuals in subgroup 3, thegroup characterized by psychiatricdisorders, had the earliest ASD di-agnoses; whether that was the casebecause their ASDs were more clearand thus diagnosed earlier, becausetheir psychiatric disorder was initially

TABLE 1 Characteristics of Each Subgroup (With 95% Confidence Intervals)

Subgroup Subgroup 1 Subgroup 2 Subgroup 3 Subgroup 4

Size 120 197 212 4316Average ICD-9 codes 150.0 (131.1–168.9) 178.6 (158.0–199.3) 103.6 (94.5–112.6) 20.8 (19.9–21.6)Proportion of boys 0.6 (0.5–0.7) 0.7 (0.6–0.8) 0.8 (0.8–0.9) 0.8 (0.8–0.8)Mean age of

diagnosis in years9.5 (8.8–10.1) 7.7 (7.1–8.3) 7.6 (7.1–8.0) 8.0 (7.9–8.1)

Autism proportion 0.8 (0.7–0.8) 0.7 (0.6–0.8) 0.6 (0.6–0.7) 0.7 (0.7–0.7)Asperger proportion 0.4 (0.3–0.5) 0.5 (0.4–0.6) 0.8 (0.7–0.9) 0.5 (0.5–0.5)PDD-NOS proportion 0.2 (0.1–0.3) 0.1 (0.1–0.2) 0.2 (0.1–0.2) 0.1 (0.1–0.1)

PDD-NOS, Pervasive Developmental Disorders, Not Otherwise Specified.

TABLE 2 Rate of Disorders in Each Subgroup (With 95% Confidence Intervals)

Outcome Subgroup 1 Subgroup 2 Subgroup 3 Subgroup 4

Seizure 77.50 (70.03–84.97) 42.13 (35.24–49.03) 33.02 (26.69–39.35) 18.16 (17.01–19.32)Psychiatric disorders 6.67 (2.20–11.13) 9.64 (5.52–13.77) 33.02 (26.69–39.35) 5.84 (5.14–6.54)GI disorders 14.17 (7.93–20.41) 24.37 (18.37–30.36) 10.85 (6.66–15.04) 3.43 (2.89–3.97)Intellectual disability 60.00 (51.23–68.77) 48.73 (41.75–55.71) 27.83 (21.80–33.86) 12.70 (11.70–13.69)Auditory disorders

and infections55.83 (46.95–64.72) 87.82 (83.25–92.38) 47.17 (40.45–53.89) 23.12 (21.87–24.38)

Cardiac disorders 30.83 (22.57–39.10) 32.99 (26.43–39.56) 24.06 (18.30–29.81) 6.93 (6.17–7.69)

ARTICLE

PEDIATRICS Volume 133, Number 1, January 2014 3

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

misdiagnosed as an ASD, or because ofthe specialties present at the hospitalcannot be determined from these data.

Characteristics of Each Subgroup

We now summarize the comorbiditiesthat were statistically significant at thea = .05 level for each subgroup afterBonferroni correction. All 802 comor-bidities were included as part of themultiple hypothesis correction, andcomplete P value tables are included inthe Supplemental Information.

To summarize the most relevant in-formation, the plots do not reveal all ofthe statistically significant comorbid-ities for each cluster, as even after

Bonferroni correction some subgroupshad tens of associated conditions. Weonly plot categories associated withspecific conditions (excluding catego-ries with names containing “other,”“nonspecific,” “unspecified,” or “symp-tom”). We also only plot comorbiditieswith elevated rates in the subgroup,have a subgroup prevalence.5%, andan interquartile range .3% (revealchanges over time).

The defining characteristic of the firstsubgroupwas seizures (Fig 4), althoughseveral comorbidities (including gas-trointestinal conditions, cerebral palsy,and disorders of the visual pathways)survived the Bonferroni correction. The

rate of convulsions rises at approxi-mately age 3 then stays steady; in-creasing diagnoses of epilepsy areseen starting at age 3 and continuingthrough age 15, likely because the con-vulsions are now being diagnosed asepilepsy.

Figure 5 shows the prevalence trajec-tory of comorbidities associated withsubgroup 2. This subgroup containedmultisystem disorders including gas-trointestinal disorders, disorders andinfections of the ear, cardiac disorders,and congenital anomalies. Individualsin this clinical phenotype had an ele-vated prevalence of seizures, psychi-atric disorders, and cerebral palsy.However, they appeared to be distinctindividuals with just one of those con-ditions as they also had the most di-agnostic codes overall (Table 1).

Subgroup 3 was characterized by psy-chiatric disorders, including episodicmood disorders, bipolar disorder, de-pression, anxiety dissociative andsomatoform disorders, conduct dis-orders, and hyperkinetic syndrome ofchildhood (Fig 6). Except for hyperki-netic syndrome of childhood, whichrose in prevalence starting at age 5,these disorders were diagnosed later,rising gradually between the ages of 5and 15. Diagnoses of anxiety-relateddisorders spiked sharply after age 11.

Finally, diagnosing ASD in an individualwith cerebral palsy or blindness can bechallenging. We recomputed the clus-tering with individuals with those di-agnoses removed. The first and thirdsubgroups (characterized by seizuresand psychiatric disorders) stayed simi-lar. The second subgroup, with indi-viduals characterized by multisystemdisorders, absorbed someof the higher-morbidity individuals from subgroup 3whenmanyofprevioushigher-morbidityindividuals were removed. Thus, theclustering patterns described aboveappear to be legitimate patterns in ASD,not just artifacts of other conditions.

FIGURE 1Prevalence of seizures and psychiatric disorders in each subgroup (with 95% confidence intervals).Subgroup 1 is characterized by seizures,whereas subgroup 3 is characterized by psychiatric disorders.Prevalence is with respect to the subgroup sizes recorded in Table 1. A, Prevalence of seizures. B,Prevalence of psychiatric disorders.

4 DOSHI-VELEZ et al

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

Validation: Separation BetweenPsychiatric and MultisystemDisorders

The 3 subgroups from our originalclustering analysis consisted of,10%of the overall sample. Thus, we couldnot validate the subgroups in a smallersample of only 496 individuals from

a different hospital (a comparison ofthe 2 hospital samples are in the Sup-plemental Information). However, wecould validate the apparent distinctionbetween individuals with only psychi-atric disorders (subgroup 3) and indi-viduals with seizures and multisystemdisorders (subgroups 1 and 2).

We hypothesized that psychiatric dis-orders would not be significantly cor-relatedwith seizures or gastrointestinaldisorders, whereas seizures and gas-trointestinal disorders would be sig-nificantly correlated. In the originalBoston Children’s Hospital sample,there was no evidence for correlationbetween psychiatric disorders andseizures (Fisher’s exact uncorrectedP = .17) or psychiatric disorders andgastrointestinal disorders (Fisher’sexact uncorrected P = .04) and strongevidence for a correlation betweenseizures and gastrointestinal dis-orders (Fisher’s exact uncorrectedP , .001).

These hypotheses were supported inthe Wake Forest individuals: aftera Bonferroni correction, we found noevidence for a correlation betweenpsychiatric disorders and seizures(Fisher’s exact uncorrected P = .13) orpsychiatric disorders andgastrointestinal

FIGURE 2Prevalence of various specific delays in development (eg, reading and coordination delays) for each of the 4 subgroups.

TABLE 3 ICD-9 Codes Contributing to at Least 10% of the Specific Delays in Development

Subgroup Specific Delays in Development (%)

Subgroup 1 315.4, Developmental coordination disorder (10.0)315.5, Mixed developmental disorder (18.1)315.8, Other specified delays in development (29.4)315.9, Unspecified delay in development (26.2)

Subgroup 2 315.31, Expressive language disorder (22.1)315.5, Mixed developmental disorder (11.3)315.8, Other specified delays in development (15.4)315.9, Unspecified delay in development (29.5)

Subgroup 3 315, Specific delays in development (11.7)315.2, Other specific developmental learning difficulties (10.4)315.31, Expressive language disorder (10.7)315.9, Unspecified delay in development (33.9)

Subgroup 4 315.31, Expressive language disorder (16.0)315.39, Other developmental speech or language disorder (10.5)315.5, Mixed developmental disorder (12.4)315.9, Unspecified delay in development (26.1)

ARTICLE

PEDIATRICS Volume 133, Number 1, January 2014 5

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

disorders (Fisher’s exact uncorrectedP = .64) but strong evidence for a cor-relation between seizures and gastro-intestinal disorders (Fisher’s exact

uncorrected P , .001). There was noevidence for a difference between thestatistics of the 2 samples (x2 simu-lated P value with Yates correction

P = .22). Our findings are consistentwith the lack of correlation betweenmedical and psychiatric comorbid-ities in Ming et al.1

DISCUSSION

The prevalence of the comorbiditiesdescribed in here are higher in the ASDpopulation than in the general pediatricpopulation, even in tertiary care cen-ters,5 and echo the differences be-tween essential and complex ASDsdescribed by Miles et al.33 Thus, al-though these co-occurrence patternsmay occur for many reasons (a dis-ease, or its treatment, may make thepatient more vulnerable for another;genetic variants may have pleiotropiceffects that include autism and othercomorbidities; and environmental in-sults may also have pleiotropic ef-fects), we do not believe these findingsare simply incidental.

Subgroup 1

Seizures,present in77.5%of individualsin subgroup 1, are 1 of the best-knowncomorbidities of autism. The mutationsresponsible for Rett syndrome21 andFragile X22 carry a much higher risk ofASD and epilepsy, but the mechanismremains poorly understood.43 The highrate of intellectual disability (60.0%) inthis subgroup is consistent with con-nections between intellectual disabil-ity, epilepsy, and ASD to the ARX(Aristaless related homeobox) gene.44

More generally, high rates of in-tellectual disability can act as a proxyfor more severe ASDs, and epilepsy isassociated with more severe ASDs.9,10

Subgroup 1 had the lowest proportionof boys, and past studies have revealedthe association between epilepsy, in-tellectual disability, and ASD to bestronger in female patients.45 Overlapbetween ASD, epilepsy, and cerebralpalsy has also been documented,46

with recent work pointing to commongenetic etiologies.47

FIGURE 3Prevalence of ASD diagnoses over time. ASD diagnoses spike dramatically starting at age 10 forindividuals for subgroupcharacterizedbyseizures (subgroup1). In contrast, individuals in thesubgroupcharacterized by psychiatric disorders (subgroup 3) have higher rates of ASD-related diagnosesstarting at age 5.

FIGURE 4Prevalence of the diagnoses characterizing subgroup 1.

6 DOSHI-VELEZ et al

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

Subgroup 2

With an intellectual disability rate of48.7%, subgroup 2 was also charac-terized by relatively severe ASDs. Sev-eral correlations found in subgroup2 have been previously reported in

autism, including cardiac and auditorydisorders,48 asthma and other auto-immune disorders,49 and congenitalanomalies involving the ear, eye, andcranial nerve.50 Konstantareas andHomatidis51 noted that the severity of

autistic features was correlated withear infections; higher rates of hearingloss have also been observed in theoverall sample.52

The complex multisystem interactionsin this group suggest a different etiol-ogy than in subgroup1.Multiple studiesof autismhavedemonstrated abnormalchemokine responses to toll-like re-ceptor ligands53,54 and abnormal natu-ral killer cell response to stimulation.55

These might contribute to an abnormalimmune response to infection andmanifest itself as increased ear infec-tions. The association between peri-partum infections and ASD may be amanifestation of the samepleiotropy.23–25

Mutations in the CaV1.2 channel56 havebeen associated with a combination ofautism, bipolar disorder, cardiac dis-orders, and immunologic disorders.However, more refinement will beneeded to understand these inter-actions.

Subgroup 3

Subgroup 3 had the highest rate ofindividuals with Asperger syndromeand the lowest rate of intellectual dis-ability (27.8%) among the 3 high-morbidity subgroups. The comorbiditiesin this group were largely psychiatricdisorders (especially anxiety). Highfunctioning children with autism oftensuffer more from anxiety than theirnormally developing counterparts,57 andcorrelations have been found betweenhigher-functioning individuals with ASDand bipolar disorder.58 A growing num-ber of studies have revealed mutationscommon to autism, attention-deficit/hyperactivity disorder, and schizophre-nia.17–20

The only nonpsychiatric comorbiditiesassociated with subgroup 3 were asth-ma and cardiac dysrhythmia.* Known

FIGURE 5Prevalence of the diagnoses characterizing subgroup 2.

FIGURE 6Prevalence of the diagnoses characterizing subgroup 3 over the first 15 years of life._

*In the PheWAS42 clustering, cardiac dysrhythmiasinclude ventricular flutter, fibrillation, and pre-mature beats; atrial fibrillation and flutter andtachycardias are not included.

ARTICLE

PEDIATRICS Volume 133, Number 1, January 2014 7

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

associations with cardiac dysrhythmiasand autism include velocardiofacialsyndrome and chromosomal disor-ders of 22q11. Velocardiofacial syn-drome is also associated with avariety of affective and neuropsychi-atric disorders besides ASD.59 Therealso exist associations between heartdisease and both anxiety60,61 and ma-jor depression,62 as well as the drugsused to treat them.63,64 The connectionsbetween the nonpsychiatric and thepsychiatric disorders suggest that psy-chiatric disorders are the main char-acteristic of this subgroup.

The 3 high-morbidity subgroups haddistinct pathophysiologies; the hetero-geneity of theASDcomorbidity spectrummirrors the heterogeneity observed ingenome-wide studies of variants asso-ciated with ASD65–67 and gene ex-pression.68 How our pathophysiologicalsubgroups map to genome-scale het-erogeneity remains unknown, but oursubgroups suggest a group structureworth investigating with these molecu-lar measures. Specifically, analyzingindividuals with ASD as a single groupmay have blurred the different etiolo-gies responsible for this heterogeneousdisease. The study presented here maysupport the exploration of the under-lying distinct pathobiologies of childrenwith ASD by providing subpopulations,which are enriched for these distinctmechanisms. If these subpopulationshave homogeneous etiological mecha-nisms, then specifically targeted thera-pies or preventive measures can beevaluated.

Limitations

Although many of the observed co-occurrence patterns are supported in

the literature, our analysis can only beconsidered preliminary because of ourrelianceon ICD-9codesand intermittenthospital visits. Originally developed forbilling, one cannot distinguish betweendiseases and symptom-complexes justfrom these codes. In particular, 364individuals in our sample had di-agnoses for ASD and cerebral palsy orblindness, which makes the diagnosisof ASD challenging. Without inspectingthe clinical notes for these patients, wecannot be sure if these individuals hadone, both, or none of these conditions.

We also have no information aboutconditions that were not diagnosed atthe particular hospital. The mean agesof the first ASD code in our study is high(between 7.6 and 9.5 years) for a dis-order that is usually diagnosed in earlychildhood. Thus, these individuals werelikely diagnosed, and treated, else-where before coming to the tertiary-care hospital. Finally, the reportedconditions are biased by the specialistsavailable at the hospital; the absence ofa condition may only indicate that thepatient sought treatment of that con-dition elsewhere.

Finally, from a technological perspec-tive, our analyses are easy to replicatein other hospital samples (as we did)because we rely only on ICD-9 codesrecorded in EHRs. However, extractingdata from health care systems doesrequire that the institution invest inextracting and standardizing the datafrom their EHR system. Although thei2b2 infrastructure41 is a free and opensource platform used by over 100 ac-ademic health centers, each of thesecenters has invested in managing thedata extraction process.69 This invest-ment is typically , 1% of the cost ofimplementing an EHR, but that portion

remains a large number in absoluteterms.

CONCLUSIONS

Three distinct medical trajectorieswere identified by unsupervised clus-tering of EHR diagnoses of individualswith ASD. Our analysis confirms theheterogeneity of the ASD, now in thelandscape of comorbidities. Each ofthese subgroups averaged more di-agnostic codes in the first 15 years oflife than the remaining overall sample.The first subgroup was characterizedby seizures, the second by multisystemdisorders, and the third by psychiatricdisorders. Each of these groupsmay point to distinct etiologies withdifferent genetic and environmentalcontributions.

Although preliminary, our study pro-vides guidance for the expensive pro-spective, longitudinal studies thatwould be needed to validate thesefindings. By providing hypotheses ofgroups to follow, our work may helptarget recruitment efforts and focusanalysis objectives. Meanwhile, furtherrefinement of these categories by usingadditional clinical and molecularcharacterizations, as well as moresophisticated time-series analysis tech-niques, will undoubtedly recover finerpatterns in the clinical trajectories ofASD.

ACKNOWLEDGMENTSThe authors thank Julie Bickel forher detailed reading and commentson the results and manuscript. Theyalso thank John Bickel and the i2b2team at Children’s Hospital Bostonfor their assistance in pulling thedata.

REFERENCES

1. Xue Ming , Brimacombe M, Chaaban J,Zimmerman-Bier B, Wagner GC. Autism

spectrum disorders: concurrent clinicaldisorders. J Child Neurol. 2008;23(1):6–13

2. Bauman ML. Medical comorbidities inautism: challenges to diagnosis and

8 DOSHI-VELEZ et al

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

treatment. Neurotherapeutics. 2010;7(3):320–327

3. Coury D. Medical treatment of autismspectrum disorders. Curr Opin Neurol.2010;23(2):131–136

4. Smith RD. Abnormal head circumference inlearning-disabled children. Dev Med ChildNeurol. 1981;23(5):626–632

5. Kohane IS, McMurry A, Weber G, et al. Theco-morbidity burden of children and youngadults with autism spectrum disorders.PLoS ONE. 2012;7(4):e33224

6. Baio J; Autism and Developmental Dis-abilities Monitoring Network SurveillanceYear 2008 Principal Investigators; Centersfor Disease Control and Prevention. Prev-alence of autism spectrum disorders—Autism and Developmental DisabilitiesMonitoring Network, 14 sites, United States,2008. MMWR Surveill Summ. 2012;61(3):1–19

7. Horvath K, Papadimitriou J, Rabsztyn A,Drachenberg C, Tildon J. Gastrointestinalabnormalities in children with autisticdisorder. J Pediatr. 1999;135(5):559–563

8. Horvath K, Perman JA. Autistic disorderand gastrointestinal disease. Curr OpinPediatr. 2002;14(5):583–587

9. Mouridsen SE, Rich B, Isager T. Epilepsy indisintegrative psychosis and infantile au-tism: a long-term validation study. Dev MedChild Neurol. 1999;41(2):110–114

10. Tuchman R, Rapin I. Epilepsy in autism.Lancet Neurol. 2002;1:352–358

11. Richdale AL, Schreck KA. Sleep problems inautism spectrum disorders: Prevalence,nature, and possible biopsychosocial etiol-ogies. Sleep Med Rev. 2009;13(6):403–411

12. Wu JY, Kuban KCK, Allred E, Shapiro F,Darras BT. Association of Duchenne mus-cular dystrophy with autism spectrumdisorder. J Child Neurol. 2005;20(10):790–795

13. Hendriksen JGM, Vles JSH. Neuropsychiat-ric disorders in males with Duchennemuscular dystrophy: frequency rate ofattention-deficit hyperactivity disorder(ADHD), autism spectrum disorder, andobsessive-compulsive disorder. J ChildNeurol. 2008;23(5):477–481

14. Hinton VJ, Cyrulnik SE, Fee RJ, Batchelder A,Kiefel JM. Association of autistic spectrumdisorders with dystrophinopathies. PediatrNeurol. 2009;41(5):339–346

15. Morgan CN, Roy M, Chance P. Psychiatriccomorbidity and medication use in autism:a community survey. Psychiatr Bulletin.2003;27:378–381

16. State MW, �Sestan N. Neuroscience. Theemerging biology of autism spectrum dis-orders. Science. 2012;337(6100):1301–1303

17. Ingason A, Rujescu D, Cichon S, et al; GROUPInvestigators. Copy number variations ofchromosome 16p13.1 region associatedwith schizophrenia. Mol Psychiatry. 2011;16(1):17–25

18. Smoller JW, Craddock N, Kendler K, et al;Cross-Disorder Group of the PsychiatricGenomics Consortium. Identification of riskloci with shared effects on five major psy-chiatric disorders: a genome-wide analysis.Lancet. 2013;381(9875):1371–1379

19. Tabarés-Seisdedos R, Rubenstein JL. Chro-mosome 8p as a potential hub for de-velopmental neuropsychiatric disorders:implications for schizophrenia, autism andcancer. Mol Psychiatry. 2009;14(6):563–589

20. Murdoch JD, State MW. Recent develop-ments in the genetics of autism spectrumdisorders. Curr Opin Genet Dev. 2013;23(3):310–315

21. Glaze DG, Percy AK, Skinner S, et al. Epilepsyand the natural history of Rett syndrome.Neurology. 2010;74(11):909–912

22. Bear MF, Huber KM, Warren ST. The mGluRtheory of fragile X mental retardation.Trends Neurosci. 2004;27(7):370–377

23. Garbett KA, Hsiao EY, Kalman S, PattersonPH, Mirnics K. Effects of maternal immuneactivation on gene expression patterns inthe fetal brain. Transl Psychiatry. 2012;2(4):e98

24. Hagberg H, Gressens P, Mallard C. In-flammation during fetal and neonatal life:implications for neurologic and neuropsy-chiatric disease in children and adults. AnnNeurol. 2012;71(4):444–457

25. Atladóttir HO, Thorsen P, Østergaard L, et al.Maternal infection requiring hospitalizationduring pregnancy and autism spectrumdisorders. J Autism Dev Disord. 2010;40(12):1423–1430

26. Jensen PB, Jensen LJ, Brunak S. Miningelectronic health records: towards betterresearch applications and clinical care. NatRev Genet. 2012;13(6):395–405

27. Roque FS, Jensen PB, Schmock H, et al.Using electronic patient records to dis-cover disease correlations and stratifypatient cohorts. PLOS Comput Biol. 2011;7(8):e1002141

28. Kohane IS. Using electronic health recordsto drive discovery in disease genomics. NatRev Genet. 2011;12(6):417–428

29. Cao H, Melton GB, Markatou M, Hripcsak G.Use abstracted patient-specific features toassist an information-theoretic measure-ment to assess similarity between medicalcases. J Biomed Inform. 2008;41(6):882–888

30. Melton GB, Parsons S, Morrison FP, RothschildAS, Markatou M, Hripcsak G. Inter-patientdistance metrics using SNOMED CT defining

relationships. J Biomed Inform. 2006;39(6):697–705

31. Hivert MF, Grant RW, Shrader P, Meigs JB.Identifying primary care patients at risk forfuture diabetes and cardiovascular diseaseusing electronic health records. BMC HealthServ Res. 2009;9:170

32. Carney CP, Jones LE. Medical comorbidity inwomen and men with bipolar disorders:a population-based controlled study. Psy-chosom Med. 2006;68(5):684–691

33. Miles JH, Takahashi TN, Bagby S, et al. Es-sential versus complex autism: definition offundamental prognostic subtypes. Am JMed Genet A. 2005;135(2):171–180

34. Wiggins LD, Robins DL, Adamson LB, BakemanR, Henrich CC. Support for a dimensionalview of autism spectrum disorders intoddlers. J Autism Dev Disord. 2012;42(2):191–200

35. Lane AE, Young RL, Baker AE, Angley MT.Sensory processing subtypes in autism:association with adaptive behavior. J Au-tism Dev Disord. 2010;40(1):112–122

36. Wing L, Gould J. Severe impairments of socialinteraction and associated abnormalities inchildren: epidemiology and classification. JAutism Dev Disord. 1979;9(1):11–29

37. Ben-Sasson A, Cermak SA, Orsmond GI,Tager-Flusberg H, Kadlec MB, Carter AS.Sensory clusters of toddlers with autismspectrum disorders: differences in affec-tive symptoms. J Child Psychol Psychiatry.2008;49(8):817–825

38. Bitsika V, Sharpley CF, Orapeleng S. An ex-ploratory analysis of the use of cognitive,adaptive and behavioural indices for clus-ter analysis of ASD subgroups. J IntellectDisabil Res. 2008;52(11):973–985

39. Hu VW, Steinberg ME. Novel clustering ofitems from the Autism Diagnostic Interview-Revised to define phenotypes within autismspectrum disorders. Autism Res. 2009;2(2):67–77

40. Sacco R, Lenti C, Saccani M, et al. Clusteranalysis of autistic patients based onprincipal pathogenetic components. AutismRes. 2012;5(2):137–147

41. Murphy SN, Weber G, Mendis M, Gainer V,Chueh HC. Serving the enterprise and be-yond with informatics for integrating bi-ology and the bedside (i2b2). J Am MedInform Assoc. 2010;17(2):124–130

42. Denny JC, Ritchie MD, Basford MA, et al.PheWAS: demonstrating the feasibility ofa phenome-wide scan to discover gene-disease associations. Bioinformatics. 2010;26(9):1205–1210

43. Tuchman R, Moshé SL, Rapin I. Convulsingtoward the pathophysiology of autism.Brain Dev. 2009;31(2):95–103

ARTICLE

PEDIATRICS Volume 133, Number 1, January 2014 9

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

44. Sherr EH. The ARX story (epilepsy, mentalretardation, autism, and cerebral malfor-mations): one gene leads to many pheno-types. Curr Opin Pediatr. 2003;15(6):567–571

45. Danielsson S, Gillberg IC, Billstedt E, GillbergC, Olsson I. Epilepsy in young adults withautism: a prospective population-based follow-up study of 120 individuals diagnosed inchildhood. Epilepsia. 2005;46(6):918–923

46. Surén P, Bakken IJ, Aase H, et al. Autismspectrum disorder, ADHD, epilepsy, andcerebral palsy in Norwegian children. Pe-diatrics. 2012;130(1). Available at: www.pe-diatrics.org/cgi/content/full/130/1/e152

47. Talkowski ME, Maussion G, Crapper L, et al.Disruption of a large intergenic noncodingRNA in subjects with neurodevelopmentaldisabilities. Am J Hum Genet. 2012;91(6):1128–1134

48. Porges SW, Macellaio M, Stanfill SD, et alRespiratory sinus arrhythmia and auditoryprocessing in autism: modifiable deficits ofan integrated social engagement system?Int J Psychophysiol. 2013;88(3):261–270

49. Becker KG. Autism, asthma, inflammation,and the hygiene hypothesis. Med Hypothe-ses. 2007;69(4):731–740

50. Miller MT, Strömland K, Ventura L, JohanssonM, Bandim JM, Gillberg C. Autism associatedwith conditions characterized by devel-opmental errors in early embryogenesis:a mini review. Int J Dev Neurosci. 2005;23(2–3):201–219

51. Konstantareas MM, Homatidis S. Ear infec-tions in autistic and normal children. JAutism Dev Disord. 1987;17(4):585–594

52. Rosenhall U, Nordin V, Sandström M, AhlsénG, Gillberg C. Autism and hearing loss. JAutism Dev Disord. 1999;29(5):349–357

53. Ashwood P, Krakowiak P, Hertz-Picciotto I,Hansen R, Pessah I, Van de Water J. Ele-vated plasma cytokines in autism spectrumdisorders provide evidence of immunedysfunction and are associated with im-paired behavioral outcome. Brain BehavImmun. 2011;25(1):40–45

54. Enstrom AM, Onore CE, Van de Water JA,Ashwood P. Differential monocyte respon-ses to TLR ligands in children with autismspectrum disorders. Brain Behav Immun.2010;24(1):64–71

55. Enstrom AM, Lit L, Onore CE, et al. Alteredgene expression and function of peripheralblood natural killer cells in children withautism. Brain Behav Immun. 2009;23(1):124–133

56. Liao P, Soong TW. CaV1.2 channelopathies:from arrhythmias to autism, bipolar dis-order, and immunodeficiency. PflugersArch. 2010;460(2):353–359

57. Gillott A, Furniss F, Walter A. Anxiety in high-functioning children with autism. Autism.2001;5(3):277–286

58. DeLong GR, Dwyer JT. Correlation of familyhistory with specific autistic subgroups:Asperger’s syndrome and bipolar affectivedisease. J Autism Dev Disord. 1988;18(4):593–600

59. Fabbro A, Rizzi E, Schneider M, Debbane M,Eliez S. Depression and anxiety disorders inchildren and adolescents with velo-cardio-facial syndrome (VCFS). Eur Child AdolescPsychiatry. 2012;21(7):379–385

60. Kubzansky LD, Kawachi I, Weiss ST, SparrowD. Anxiety and coronary heart disease:a synthesis of epidemiological, psychologi-cal, and experimental evidence. Ann BehavMed. 1998;20(2):47–58

61. Watkins LL, Grossman P, Krishnan R, SherwoodA. Anxiety and vagal control of heart rate.Psychosom Med. 1998;60(4):498–502

62. Stapelberg NJ, Hamilton-Craig I, NeumannDL, Shum DH, McConnell H. Mind and heart:heart rate variability in major depressivedisorder and coronary heart disease–a re-view and recommendations. Aust N Z JPsychiatry. 2012;46(10):946–957

63. Witchel HJ, Hancox JC, Nutt DJ. Psychotro-pic drugs, cardiac arrhythmia, and suddendeath. J Clin Psychopharmacol. 2003;23(1):58–77

64. Taylor D. Antidepressant drugs and car-diovascular pathology: a clinical overviewof effectiveness and safety. Acta PsychiatrScand. 2008;118(6):434–442

65. Liu L, Sabo A, Neale BM, et al Analysis ofrare, exonic variation amongst subjectswith autism spectrum disorders and pop-ulation controls. PLoS Genet. 2013;9(4):e1003443

66. Yu TW, Chahrour MH, Coulter ME, et al.Using whole-exome sequencing to identifyinherited causes of autism. Neuron. 2013;77(2):259–273

67. Michaelson JJ, Shi Y, Gujral M, et al. Whole-genome sequencing in autism identifies hotspots for de novo germline mutation. Cell.2012;151(7):1431–1442

68. Kong SW, Collins CD, Shimizu-Motohashi Y,et al Characteristics and predictive value ofblood transcriptome signature in maleswith autism spectrum disorders. PLoS ONE.2012;7(12):e49475

69. Masys DR, Harris PA, Fearn PA, Kohane IS.Designing a public square for researchcomputing. Sci Transl Med. 2012;4(149):49fs32

10 DOSHI-VELEZ et al

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

originally published online December 9, 2013; Pediatrics Finale Doshi-Velez, Yaorong Ge and Isaac Kohane

Record Time-Series AnalysisComorbidity Clusters in Autism Spectrum Disorders: An Electronic Health

ServicesUpdated Information &

013-0819http://pediatrics.aappublications.org/content/early/2013/12/03/peds.2including high resolution figures, can be found at:

Supplementary Material

013-0819.DCSupplementalhttp://pediatrics.aappublications.org/content/suppl/2013/12/03/peds.2Supplementary material can be found at:

Permissions & Licensing

https://shop.aap.org/licensing-permissions/in its entirety can be found online at: Information about reproducing this article in parts (figures, tables) or

Reprintshttp://classic.pediatrics.aappublications.org/content/reprintsInformation about ordering reprints can be found online:

ISSN: . 60007. Copyright © 2013 by the American Academy of Pediatrics. All rights reserved. Print American Academy of Pediatrics, 141 Northwest Point Boulevard, Elk Grove Village, Illinois,has been published continuously since . Pediatrics is owned, published, and trademarked by the Pediatrics is the official journal of the American Academy of Pediatrics. A monthly publication, it

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from

originally published online December 9, 2013; Pediatrics Finale Doshi-Velez, Yaorong Ge and Isaac Kohane

Record Time-Series AnalysisComorbidity Clusters in Autism Spectrum Disorders: An Electronic Health

http://pediatrics.aappublications.org/content/early/2013/12/03/peds.2013-0819located on the World Wide Web at:

The online version of this article, along with updated information and services, is

ISSN: . 60007. Copyright © 2013 by the American Academy of Pediatrics. All rights reserved. Print American Academy of Pediatrics, 141 Northwest Point Boulevard, Elk Grove Village, Illinois,has been published continuously since . Pediatrics is owned, published, and trademarked by the Pediatrics is the official journal of the American Academy of Pediatrics. A monthly publication, it

by guest on January 9, 2018http://pediatrics.aappublications.org/Downloaded from