Data-driven dissection of the fever effect in autism spectrum disorder

Some individuals with autism spectrum disorder (ASD) demonstrate marked behavioral improvements during febrile episodes, in what is perhaps the only present-day means of modulating the core ASD phenotype. Understanding the nature of this so-called fever effect is therefore essential for leveraging this natural temporary relief of symptoms to a sustained efficacious intervention. Toward this goal, we used machine learning to analyze the rich clinical data of the Simons Simplex Collection, in which one out of every six children with ASD was reported to improve during febrile episodes, across multiple ASD domains. Reported behavioral improvements during febrile episodes were associated with maternal infection in pregnancy (OR = 1.7, 95% CI = [1.42, 2.03], P = 4.24 (cid:1) 10 (cid:3) 4 ) and gastrointestinal (GI) dysfunction (OR = 1.46, 95% CI = [1.15, 1.81], P = 1.94 (cid:1) 10 (cid:3) 3 ). Family members of children reported to improve when febrile have an increased prevalence of autoimmune disorders (OR = 1.43, 95% CI = [1.23, 1.67], P = 3.0 (cid:1) 10 (cid:3) 6 ), language disorders (OR = 1.63, 95% CI = [1.29, 2.04], P = 2.5 (cid:1) 10 (cid:3) 5 ), and neuropsychiatric disorders (OR = 1.59, 95% CI = [1.34, 1.89], P < 1 (cid:1) 10 (cid:3) 6 ). Since both GI abnormalities and maternal immune activation have been linked to ASD via proinflammatory cytokines, these results might suggest a possible involvement of immune dysregulation in the fever effect, consistent with findings in mouse models. This work advances our understanding of the fever-responsive ASD subtype and motivates the future studies to directly test the link between proinflammatory cytokines and behavioral modifications in individuals with ASD


INTRODUCTION
During the first half of the twentieth century, artificial fever therapy (pyrotherapy) was widely used to treat an array of neuropsychiatric disorders (Epstein, 1936;Raju, 2006). In the second half of the century, caregivers and parents of children with autism have anecdotally reported marked behavioral improvements during febrile episodes (Cotterill, 1985;Sullivan, 1980). Today, ASD remains one of the most debilitating childhood disorders: It affects 2.27% of the US population (Maenner et al., 2021), and requires chronic management (Harrington & Allen, 2014). Understanding the so-called fever effect in ASD could lead to precision approaches to therapy, based on its underlying neurobiological mechanisms.
Toward this goal, a prospective study of behavioral changes associated with fever in children with ASD was the first to document fewer aberrant behaviors in 30 children with ASD during febrile episodes as compared to 30 matched afebrile children (Curran et al., 2007). More recently, Byrne and colleagues prospectively monitored 141 children with ASD and 103 typically developing controls for a period of 3 months. Of these, 50 children with ASD had a fever during the reporting period and three reported behavioral improvements when febrile (6%, 95% CI = [1.26%, 16.55%]), as compared to no such reported improvements among controls (Byrne et al., 2022). Moreover, a large-scale investigation of parental reports in the Simons simplex collection (SSC) found that one in six children with ASD reportedly improved when he or she had a fever (16.7%, 95% CI = [15.2%, 18.3%]), across multiple ASD domains, including temper, communication, social interaction, repetitive behavior, and cognition (Grzadzinski et al., 2017). Children reported to improve when febrile tend to have more severe ASD symptoms, such as more repetitive behaviors, and lower non-verbal cognitive skills. However, that study focused on a selected set of variables from the rich SSC dataset, mostly those related to the child's behavior and demographics. Additional characteristics of fever responsive children remain to be identified. A comprehensive profiling of these children and their families could shed light on the neurobiological mechanisms underlying the fever effect.
Recent studies in animal models of ASD have shown that interleukin-17a (IL-17a), a pro-inflammatory cytokine often secreted during fever, but not fever per se, rescues social deficits in autistic mice by binding its receptor (IL-17Ra) on neurons of the primary somatosensory cortex dysgranular zone (S1DZ). Importantly, IL-17a produced peripherally as part of a typical immune insult was shown to restore sociability in mouse models of ASD caused by maternal immune activation (MIA), but not in other monogenic models of ASD (Reed et al., 2020). The authors attribute the increased exposure of MIA mice to IL-17a during fetal development to increased neuronal sensitivity for IL-17a after birth (Gloria Choi, 2022). Accordingly, direct delivery of IL-17a to the S1DZ of monogenic models of ASD resulted in the same behavioral improvements, as measured by the three-chamber social approach assay and the reciprocal social interaction assay (Reed et al., 2020). Thus, the ability of S1DZ neurons to sense IL-17a might distinguish animals whose sociability improves during systemic inflammation from those who do not. IL-17a signaling was also shown to mediate gastrointestinal inflammation in MIA models of ASD, and overall prime the mouse immune system via epigenetic alterations of naïve CD4+ T cells . This ancient cytokine was also shown to act as a neuromodulator in Caenorhabditis elegans, regulating worm behavior (Chen et al., 2017;Flynn et al., 2020). Yet, if and how these findings may translate to humans remains to be elucidated. Immune dysregulation is a well-established etiology for a subset of individuals with ASD (Croen et al., 2019;Ramaswami et al., 2020;Robinson-Agramonte et al., 2022), and increasing levels of proinflammatory cytokines have been associated with GI dysfunction in ASD (Ashwood & Wakefield, 2006;Jyonouchi et al., 2011;Samsam et al., 2014), as well as with ASD severity (Ashwood et al., 2011;Masi et al., 2017). IL-17a receptors are expressed throughout the developing and adult human brain (Fujitani et al., 2022), and plasma IL-17a levels were shown to correlate with cognitive function in individuals with Parkinson disease (Green et al., 2019) and multiple sclerosis (Trenova et al., 2018).
It should be noted that the accumulating knowledge of the fever effect remains limited as most studies were conducted on animal models of ASD (Choi et al., 2016;Reed et al., 2020), relied on modest sample sizes, and, when applied to humans were based on parental reports and/or retrospective memory (Byrne et al., 2022;Curran et al., 2007;Grzadzinski et al., 2017). These and other factors, including the high ASD heterogeneity and different study designs have contributed to gaps in our understanding of the proportion of autistic children who may improve with fever.
Identifying the clinical characteristics of individuals with ASD reported to improve when febrile will advance our understanding of the neurobiological mechanisms underlying behavioral modifications in ASD. We therefore used various machine learning approaches to dissect the rich phenotypic data of the SSC (Fischbach & Lord, 2010) and comparatively characterize the fever responsive subgroup of individuals with ASD, in an unbiased, data-driven fashion.

Patients and datasets used
We mined the SSC data repository (version 15), which includes comprehensive clinical, behavioral, genomic, and demographic data from 2951 families with one child diagnosed with ASD (Fischbach & Lord, 2010). As previously described (see https://www.sfari.org/resource/ simons-simplex-collection/), the SSC inclusion criteria are proband age between 4 and 17.9 years, Autism Diagnostic Interview-Revised (ADI-R) and Autism Diagnostic Observation Schedule (ADOS) based diagnoses, and minimal nonverbal cognitive abilities as measured by the Differential Ability Scales-II (DAS-II), Wechsler Intelligence Scale for Children (WISC-IV), or Wechsler Abbreviated Scale of Intelligence (WASI). Exclusion criteria included pregnancy or birth complications, nutritional or psychological deprivation, and family members with ASD, mental retardation, or schizophrenia. Of 2951 SSC families, we excluded those who did not report on their child's behavior during fever, resulting in a final sample of 2253 families. Informed consent was obtained and approved by local Institutional Review Boards at each of the 12 data collection sites included in the SSC (https://www.sfari. org/resource/simons-simplex-collection-sites/).
Consequently, the Boston Children's Hospital Institutional Review Board has determined that this particular research qualifies as exempt from the requirements of human subjects protection regulations.

Feature selection
With a goal of pointing to likely underlying neurobiological mechanisms, we focused on medical variables of the SSC Phenotype Dataset, as well as those considered core and commonly used. The full description and definitions of all phenotypic data can be downloaded from https:// www.sfari.org/resource/simons-simplex-collection/. In all, 404 variables were selected from 29 SSC tables based on these criteria (Supplementary Table 1). Of these, 185 sparse and/or closely related variables were aggregated into 29 summary statistics (Supplementary Table 2). Then 44 variables were removed due to having >30% missing values, and 20 others were removed for having a constant value across the entire cohort. Supplementary Figure 1 depicts this feature selection process. The SSC table column of Supplementary Table 1 details the source of each variable, and identifies those core and commonly used as variables included in the ssc_core_descriptive and ssc_commonly_used tables, respectively. Detailed information about how constructs were measured, and which measures were used is available at https://simonsfoundation.s3.amazonaws.com/share/SSC_ DataDictionary.xls.zip

Continuous variable clustering
Pairwise correlations were calculated between 82 continuous variables. Variables were represented as nodes and edges joined two nodes if their R 2 ≥ 0.7. Clusters were defined as connected components in this graph. A total of 11 multivariate clusters and 31 single feature clusters were identified (Supplementary Figure 2 and Table 3).

Categorical variable clustering
Pairwise Fisher's exact tests were calculated between 244 categorical variables, using Monte Carlo simulations with B = 1 Â 10 7 where needed. Variables were represented as nodes and edges joined two nodes if their Fisher's exact association P-value was ≤9 Â 10 À6 . This specific P-value threshold was selected as the Benjamini-Hochberg corrected P-value at the jump point in the number of clusters as a function of the P-value (Supplementary Figure 3A). Clusters were defined as communities of information flow (Rosvall & Bergstrom, 2008) in this variable association graph. A total of 15 multivariate clusters and 83 single feature clusters were identified using R's igraph package (Csardi & Nepusz, 2006) (Supplementary Figure 3B and Table 4).

Characterizing the population of fever responders
Variable clusters described above were tested against a binary target of parental response to whether their child seems to show any improvement in symptoms of autism when s/he has a fever? Children whose parents answered "yes" were labeled as fever responders, those whose parents replied "no" were labeled non-responders, and those whose parents were unsure were ignored. Mann-Whitney U tests were used to assess differences in continuous clusters between responders and non-responders, while Fisher's exact tests were used for inter-group comparisons of categorical clusters. We used bootstrap validation to stabilize the analysis by repeatedly resampling (n = 10,000) equally sized populations from both groups and calculating bootstrapped statistics. Throughout the article, we report median P-values and odds ratios (ORs), as well as 95% confidence intervals (CIs) of the bootstrap samples.

Domains of improvement analysis
For responders, parental reports included a yes/no answer to observing behavioral changes in the following ASD domains: cognition, social skills, communication, temper, and repetitive behavior. Fisher's exact tests were used to examine dependencies between domains.
Network structure of fever response characteristics To identify clinical features indirectly associated with fever response, the analysis described under "Characterizing the population of responders" was repeated for the strongest medical predictors of fever response, as detailed in Supplementary Table 4. Specifically, the target variable was changed from fever response to (A) GI dysfunction (the centroid of Cluster 1 in Supplementary Table 4), and (B) maternal infection in pregnancy (the centroid of Cluster 3 in Supplementary Table 4). The most associated features were then joined to a network based on their ORs.
Assessing the relative importance of maternal infection during pregnancy and GI dysfunction To evaluate the relations between maternal infection during pregnancy and GI disorders with fever response, while controlling for the main cognitive and behavioral predictors of fever response, we fitted a logistic regression model of fever response as a function of all major fever response predictors. As another indication of the relative importance of these features, we further fitted two additional models, each time omitting one of the two variables (maternal infection in pregnancy, sum severity GI disorders) and examining the resulting model fit, as compared to the full model. The "anova" function (with test = "chisq," stats package [R Core Team, 2021]) was used to compare models.

Multivariate predictive power assessment
Continuous and categorical variable clusters found to be related to fever response were integrated into a multivariate classifier using forward feature selection in a support vector machine (SVM), random forest (RF), and generalized linear model (GLM). A subset of 1340 individuals (235 responders and 1105 non-responders) that had complete data in all 20 significant variable clusters was included in this analysis. The relative frequency of responders in this sample is consistent with that of the entire SSC. No imputation was used in this analysis. The R packages kernlab (Karatzoglou et al., 2004), random-Forest (Liaw & Wiener, 2002), and stats (R Core Team, 2021) were used for model fitting. Repeated (n=100) 10-fold cross validation was used to assess the models' area under the receiver operator characteristics curve (AUC).

Assessing the relative importance of maternal awareness
We examined the importance of maternal social cognition, as measured by the SRS, in reporting a proband's fever response relative to other variables in the multivariate classifier described above. The SRS-defined maternal social cognition variable was available for all but three families. No missing data imputation was performed. A GLM and RF were used to rank variable importance and compare the predictive power of an integrative model fit with and without the maternal social cognition variable. The R packages stats (R Core Team, 2021), randomForest (Liaw & Wiener, 2002), and ROCR (Sing et al., 2005) were used for GLM training, RF modeling, and calculating the AUC, respectively.
Post hoc comparative analysis of sensory impairments in fever responders We compared the relative frequency of score 2, "definite abnormal behavior", in ADI-R question 71, "Unusual Sensory Interests-ever" between fever responders and non-responders. Fisher's exact test was used to test for the association between score 2 in ADI-R question 71 and fever response.

Multiple testing correction
All P values were corrected for multiple comparisons using the Benjamini-Hochberg procedure (Benjamini & Hochberg, 1995). The false discovery rate (FDR) of this study was controlled at α = 0.05.

RESULTS
Considering 404 relevant clinical characteristics of the SSC Phenotype Dataset (Supplementary Table 1 and Methods), we identified 17 clusters of continuous variables significantly correlated with fever response (Supplementary Table 3), and three clusters of categorical variables significantly associated with fever response (Supplementary Table 4). We refer to each cluster with the name of its centroid variable. The most salient differences between fever responders and non-responders were behavioral and cognitive, all previously reported by Grzadzinski et al. in their analysis of the SSC phenotypic data (Supplementary Table 3 and Grzadzinski et al., 2017). Here, we focus on presenting findings of our unsupervised data-driven analyses that were not previously reported.

Fever response is strongly related to GI dysfunction in children with ASD
Children reported to improve during febrile episodes were more likely to be reported as suffering from various GI dysfunctions, including diarrhea, bloating, and severe abdominal pain (Supplementary Figure 5, OR = 1.46, 95% CI = [1.15, 1.81], P = 1.94 Â 10 À3 ). The more severe the overall GI symptoms, the greater the chance for observed behavioral improvement during febrile episodes (Figure 2b, Wilcoxon P = 8.81 Â 10 À18 ). This tight correlation is linked to the behavioral and cognitive differences between responders and non-responders, as well as to maternal infections in pregnancy. Specifically, controlling for behavioral and cognitive features, as well as maternal infections in pregnancy, GI symptoms severity per se was only marginally predictive of fever response in a logistic regression model (P = 0.076, Supplementary Table 5), and omitting it from the model reduced the model fit with P = 0.081.

Network analysis identifies relationships between fever response, maternal infection in pregnancy, and GI dysfunction
To clinically characterize the ASD subtype of children who responded to fever, we also examined clinical features indirectly associated with fever response. This analysis revealed that autoimmune disorders in the family, neuropsychiatric disorders in the family, and language disorders in the family are all characteristics of fever response, linked via GI dysfunction and maternal infection during pregnancy (Figure 3).

Maternal social cognition confounds fever response characterization
Maternal social cognition scores, as measured by the SRS and reported by spouses, were lower for mothers of children reported to improve when febrile, reflecting a greater maternal ability to interpret social cues once they are picked up (Supplementary Figure  6A, P = 3.94 Â 10 À3 ). No relationship was detected between paternal social cognition and a proband's fever response (Supplementary Figure 6B, P = 0.17). Since SSC mothers tend to be those responsible for most parental reports, this finding suggests that maternal social abilities represent a confounder in this analysis, which future study designs should minimize.

Multivariate analysis demonstrates that further fever effect determinants remain to be discovered
We used multivariate modeling to understand the overall and relative contributions of the identified fever effect determinants. Forward feature selection was first applied to identify the optimal combination of clinical characteristics most informative of fever response (Supplementary Figure 7). Using repeated 10x cross validation, we found that logistic regression with nine variables predicted fever response with a mean AUC of 0.696 (0.056) (Supplementary Figure 7). Variable ranking identified the most informative determinants to be verbal IQ, ADI-R verbal communication score, and ADI-R repetitive and restricted behavior (RRB) score, all previously reported by Grzadzinski et al. (Supplementary Figure 8). We used two complementary approaches to assess the F I G U R E 2 Predictors of fever response. (a) Maternal infection in pregnancy is strongly associated with fever response in the proband (OR = 1.7, 95% CI = [1.42, 2.03], P = 4.24 Â 10 À4 ). The plot shows the percent of individuals reported to improve when febrile out of all those with answers to questions about maternal infections in pregnancy. (b) Fever responders tend to be children with ASD and more severe GI comorbidities (Wilcoxon P = 8.81 Â 10 À18 ).
impact of the maternal SRS social cognition confounder. In the first, we excluded this variable from the optimal model and compared the model's performance with and without the maternal social cognition variable. Compared to the full model's mean AUC of 0.696 (0.056), the maternal SRS social cognition-null model performed with a mean AUC of 0.678 (0.059), suggesting that although maternal cognition confounds our analysis, its overall contribution to the fever effect characterization is modest (Supplementary Figure 9). In the second approach, we matched responders and non-responders based on their maternal SRS social cognition scores, and tested the predictive power of the optimal multivariate model on these subpopulations. The resulting mean AUC of 0.679 (0.061) (Supplementary Figure 10) suggests that further fever effect determinants remain to be discovered.

DISCUSSION
This data-driven analysis of the fever effect in ASD provides a comprehensive characterization of the feverresponsive ASD subtype. The previous study analyzing the same data focused mainly on behavioral and cognitive characteristics of fever responders, highlighting the role of repetitive and restrictive behaviors and lower non-verbal IQ as their most distinctive features (Grzadzinski et al., 2017). The machine learning approaches utilized in our study represent complementary analyses that support and extend their findings, focusing on medical characteristics of probands and their family members. While the nonbehavioral differences are more modest than the behavioral ones, our network analysis suggests that they could hold the key to understanding how the fever effect is linked with other neuropsychiatric, autoimmune, and language disorders, and point to likely neurobiological mechanisms.
Our work is consistent with recent studies in ASD animal models exposed to maternal immune activation (MIA) in utero. Specifically, findings in mice show that MIA causes elevated production of proinflammatory cytokines such as IL-17a (Choi et al., 2016;Kalish et al., 2021;Reed et al., 2020), which cross the placenta and affect mRNA translation in the developing fetal brain, increasing the risk for neurodevelopmental disorders such as ASD (Kalish et al., 2021;Reed et al., 2020). Moreover, IL-17a was shown to alter the maternal gut microbiota, which renders the immune system of the offspring more susceptible to GI inflammatory responses Kim et al., 2022). Importantly, MIA models of ASD show improvement in social behaviors during lipopolysaccharide (LPS)-induced inflammation, mediated by elevated IL-17a production (Reed et al., 2020). Accordingly, our findings provide the first evidence for a subset of human subjects with ASD exposed to maternal immune activation in utero, with F I G U R E 3 Network structure of the clinical characteristics of children with ASD responding to fever. Shown are clinical features directly and indirectly associated with fever response, their ORs, 95% CIs, and association P values. comorbid GI dysfunction, and improved behaviors during febrile episodes. Our findings suggest that proinflammatory cytokines might underly the reported beneficial effects of fever in this subgroup.
Moreover, the association of familial autoimmune, neuropsychiatric, and language disorders with this subtype provides further support for the involvement of proinflammatory cytokines, since these disorders share a well-documented immune dysregulation etiology (Cerri et al., 2017;Estes & McAllister, 2016;Hunter & Jones, 2015;Knuesel et al., 2014;Tabarkiewicz et al., 2015). One implication of our findings is that proinflammatory cytokines might also have beneficial effects in these disorders, consistent with findings in MIA mice. Indeed, several anecdotal reports describe symptom amelioration in some individuals with attention deficit hyperactivity disorder (ADHD) (Marner, 2018), schizophrenia (Zuschlag et al., 2016), and bipolar disorder (BP) (Setsaas & Vaaler, 2014) during febrile episodes. Further research should characterize the fever effect in these neurodevelopmental disorders. Though some changes in behavior at times of illness are expected (e.g., higher lethargy, reduced aggressiveness), improvements such as enhanced communication skills, less repetitive behaviors, or cognitive improvements are surprising. For ASD, a common neurodevelopmental disorder with no pharmacological treatment options, this natural remedy represents one of few leads toward therapy.
Our study has several limitations. Its main limitation stems from its source of data, namely retrospective parent reports. A reliance on retrospective memory introduces several biases to the data. While major medical events and demographic features during pregnancy (such as time of pregnancy) are reported accurately, other events, including maternal infection, are typically underestimated (Buka et al., 2004;Voldsgaard et al., 2002). Parental reports might introduce additional bias, as evidenced by the importance of the mother's social cognition, reported by spouses, and revealed in an unsupervised data-driven approach. Besides that, other parental biases not directly captured in the data might have affected our results. These include the amount of time the parents spend with their child, parental tendency to track symptoms and be alerted to changes, and the frequency of febrile episodes in the child.
Another limitation of this study is that by basing our analyses on a simple "yes/no" target we ignore the degree of improvement, the possibility of an opposite effect (i.e., worse behavior), the fever effect's dynamics throughout development, and more precise diagnoses reflecting the cause of fever during which improvement was observed. Additionally, other factors may influence a child's behavior when he or she are sick, including increased parental or caregiver stress, increased attention, changes in sleep patterns, and changes in nutrition. Furthermore, it should be noted that our study lacks a control group and thus we were unable to compare parental reports on improvements in cognition or social communication that might have occurred during fever in typically developing children. However, a recent prospective study by Byrne et al. (2022) that used such a control group, did not find any consistent social or behavioral improvements during fever in typically developing children.
Notably, the Byrne et al.'s study identified behavioral improvements during fever in 6% (95% CI = [1.26%, 16.55%]) of children with ASD, as compared to 16.7% (95% CI = [15.2%, 18.3%]) of children with ASD reported to improve when febrile in the SSC cohort analyzed in the present study. Several factors may have contributed to different prevalence estimates in the two studies. First, the observation period in the Byrne et al.'s study was 3 months, while that of the present study was 8.88 years (±3.5). While the Byrne et al.'s study was a prospective one, the present study relied on parental reports of any improvement in ASD symptoms during any febrile episode throughout the child's lifetime. Since studies in animal models suggest that behavioral improvements during febrile episodes are driven by the specific immune response underlying the observed fever (Reed et al., 2020), a longer observation period provides more chances for more febrile episodes mediated by such immune mechanisms. It is also possible that the reliance on retrospective memory in the present study inflates fever response rate estimates, as compared to the better controlled, prospective Byrne et al.'s study. Another methodological difference between the two studies that might affect prevalence estimates is that the Byrne et al.'s study monitored fever response in 50 children with ASD, while ours analyzed parental reports on 2253 probands. In doing so, the Byrne et al.'s study might have under-sampled the high-ASD heterogeneity, potentially under-representing fever responsive children. Moreover, different inclusion and exclusion criteria of the two cohorts might have contributed to different fever response rate estimates. For example, children with severe sensory impairment were excluded from the Byrne et al.'s cohort, yet a post hoc comparative analysis of sensory impairment among fever responsive and non-responsive SSC probands revealed that fever response was strongly associated with unusual sensory interests as measured by the ADI-R (OR = 1.715, 95% CI = [1.358, 2.163], P = 3.0 Â 10 À6 ).
Follow-up prospective studies could resolve our study's limitations by (A) minimizing parental reporting biases, for example, with the help of technology and tracking behavioral variability across the spectrum, (B) dissecting the immune processes accompanying behavioral changes, (C) incorporating electronic health record data for a more precise characterization of the fever-responsive ASD subtype, (D) fine-mapping the behavioral domains that do and do not vary in response to immune insults. For example, a smartphone app could be used to track behavioral changes and comorbid conditions in real time.

CONCLUSIONS
An emerging subtype of fever-responsive ASD may be characterized by GI abnormalities, maternal infection in pregnancy, and familial autoimmune and neuropsychiatric disorders. Our work suggests that the fever effect might be shared across neurodevelopmental disorders, implying that similar analyses would be informative for dissecting ADHD, schizophrenia, and BP. Our findings suggest that future studies of circulating proinflammatory cytokine dynamics and their relations to fever and behavioral changes in human subjects could pave the way toward the development of targeted treatment options for this ASD subtype and other immunemediated neurodevelopmental disorders.