Autism spectrum disorder (ASD) is one of the neurodevelopmental disorders that is characterized by two core symptoms: difficulties with social communication and interaction and the presence of repetitive and restricted behaviors or interests (American Psychiatric Association, 2013). Parents of ASD patients were found to experience greater parenting stress than typically developing individuals and even other disabilities (Hayes and Watson, 2013). It is an important issue because high-level parenting stress is associated with the lower effectiveness of parent-mediated intervention (Osborne et al., 2008). Therefore, helping stressed parents can be beneficial in improving the outcome of treatment for ASD patients.

Numerous studies have explored associated factors for parenting stress in caregivers of ASD patients, and personality traits and mood problems of caregivers (Falk et al., 2014; Leonardi et al., 2021), ASD core symptoms (Miranda et al., 2019; Scibelli et al., 2021), and behavioral problems of ASD patients (Yorke et al., 2018; Miranda et al., 2019; Siu et al., 2019; Scibelli et al., 2021) were found to be significantly linked with parenting stress. Recently, associated factors for each dimension of parenting stress (parental distress, parent-child dysfunctional interaction, and difficult child) were investigated (Mello et al., 2022). However, previous studies have utilized different combinations of study variables and statistical tests, leading to inconsistent results and difficult interpretations (Voliovitch et al., 2021; Mello et al., 2022). Moreover, some studies did not address the study variables at the same level; for example, the total score was used for ASD core symptoms, but subscales for behavioral problems (Siu et al., 2019; Mello et al., 2022). Lastly, none of them have attempted to apply machine learning (ML) methods which offer distinct advantages over traditional approaches since ML can handle multi-dimensional and non-linear relationships (Schwalbe and Wahl, 2020).

Herein, our study aimed at identifying predictive features for parenting stress (parental distress, parent-child dysfunctional interaction, difficult child, and total parenting stress) in caregivers of ASD patients by developing explainable ML models. Additionally, we included only subscales of the Minnesota Multiphasic Personality Inventory-2 (MMPI-2) for caregivers’ psychopathology (Graham, 1990; Han et al., 2006), Social Responsiveness Scale-2 (SRS-2) for ASD core symptoms (Constantino, 2012; Chun et al., 2021), Child Behavior Checklist (CBCL) for behavioral problems (Han and Yoo, 1995; Achenbach, 1999), and other additional features as model features. We expected that the identified predictive features would help alert clinicians to whether a caregiver is at a high risk of severe parenting stress and provide timely interventions to stressful parents, which would eventually enhance the treatment outcome for ASD patients.

We followed the STROBE guideline (Supplementary material, pp. 6–7) (von Elm et al., 2007). The present study was approved by the Institutional Review Board of the Severance Hospital of Yonsei University, Seoul, the Republic of Korea. Informed consent was waived since we used retrospective and deidentified patient data (IRB number: 4-2022-0803). The overall process of ML models is displayed in Figure 1.

We evaluated the ML models predicting severe parenting stress and its components (parental distress, parent-child dysfunctional interaction, and difficult child) in caregivers of ASD patients and investigated major predictors. Our key findings were that our ML models could predict severe parental distress, parent-child dysfunctional interaction, difficult child, and total parenting stress with AUC values greater than 0.80. Moreover, we also identified major predictors for each outcome of interest by utilizing explainable ML models, which provided valuable insights into the underlying factors contributing to severe parenting stress in caregivers of ASD patients.

Parental distress measures a parent’s experiences of their role as parents (Abidin, 1990). Among the top 10 predictors for parental distress, seven were associated with the personality traits of caregivers [depression (code 2), schizophrenia (code 8), psychopathy (code 4), psychasthenia (code 7), paranoia (code 6), social introversion (code 0), and hypomania (code 9)], which means that caregivers’ perceived hardship related to the role as parents may be determined primarily by their psychopathology. However, an observational study using regression analysis reported that ASD patients’ emotional problems (regression coefficient = 0.31) may also play a significant role in parental distress (Mello et al., 2022). Since our main ML models using the overall sample only included common subscales of CBCL between 1.5–5 and 6–18, and thereby scores of emotional problems were excluded, the potential impact of patients’ behavioral problems on caregivers’ parental distress should not be ignored. Indeed, our subgroup analysis that used CBCL 6–18 sample showed that the patients’ behavioral problems were also essential predictors in predicting parental distress in caregivers of ASD patients. Nevertheless, when considering that most previous studies only utilized ASD patient factors as associated/predictive factors for parental distress (Scibelli et al., 2021; Mello et al., 2022), our study provided a new insight into the understanding of parental distress by utilizing explainable ML models with model features related to caregivers’ psychopathology. However, interpretation needs caution since MMPI-2 clinical scales should not be independently addressed (Levak et al., 2012). A comprehensive approach for significant MMPI-2 clinical scales might be appropriate. For example, we may expect that a caregiver of 2-4-8 code type (the top 3 predictors for parental distress) would experience substantial difficulties in their role as parents as to the vulnerability to substance abuse, poor impulse control, emotional dysregulation, posttraumatic stress disorder symptoms, thought disorder, and borderline personality disorder, which might represent moderate to severe psychopathology and require major psychiatric interventions (Archer et al., 1995; Bell-Pringle et al., 1997; Donovan et al., 1998).

Parent-child dysfunctional interaction measures parents’ feelings about the interaction with their child with ASD (Abidin, 1990). Interestingly, our findings implied that behavioral problems of children (withdrawn, aggressive behavior, and attention problems) contributed more to caregivers’ negative feelings on their interaction with the children compared to ASD core symptoms, even though the latter ones seemed to be more directly associated with their interaction. Furthermore, a cross-sectional study observed that the correlation coefficient (r) between total ASD core symptoms and parent-child dysfunctional interaction (0.305) was attenuated when considering its individual subscales. Specifically, the correlations with parent-child dysfunctional interaction were estimated to be 0.184 for reciprocal social interaction, 0.212 for social communication, and 0.288 for repetitive and restricted behaviors (Scibelli et al., 2021). However, a separate study that employed a regression analysis highlighted the significance of ASD core symptoms on parent-child dysfunctional interaction (ASD core symptom severity: regression coefficient = 0.48, p < 0.001) (Mello et al., 2022). This indicates that the impact of ASD core symptoms should not be underestimated and remains a significant concern. In fact, two of the top 5 predictors were related to ASD core symptoms (social communication and cognition) in the subgroup analysis using the CBCL 1.5–5 sample. Taken together, our results suggested that both ASD core symptoms and behavioral problems of ASD patients are important in predicting parent-child dysfunctional interaction.

Difficult child measures a parent’s perception of whether the child is easy or difficult to nurture (Abidin, 1990). Our models found that ASD patients’ behavioral problems (aggressive behavior, anxious/depressed, other problems, attention problems, and withdrawn), caregivers’ psychopathology [social introversion (code 0) and depression (code 2)], and ASD patients’ core symptoms (autistic mannerisms and social cognition) mainly contributed to the caregivers’ perception that their child was difficult to care. These results are consistent with the previous studies that have reported associated/predictive factors for difficult child in parents of ASD patients. Scibelli et al. found that difficult child (as measured by PSI-SF) was significantly correlated with ASD core symptoms (total symptoms: r = 0.409; repetitive and restricted behavior: r = 0.409) and behavioral problems (social problems: r = 0.485; thought problems: r = 0.671; attention problems: r = 0.502; rule-breaking behavior: r = 0.498; aggressive behavior: r = 0.555) in ASD without cognitive impairment (Scibelli et al., 2021). Mello et al. (2022) have also reported significant impacts of ASD core symptoms (regression coefficient = 0.46) and aggressive behavior (regression coefficient = 0.36) on difficult child. Together with our findings, this suggested that the severity of core symptoms and behavioral problems seemed to directly contribute to the difficulties faced by caregivers in their upbringing responsibilities. However, our study was the first to report that caregivers’ personality traits were also major predictors for difficult child, even outranking ASD core symptoms. That is, the caregivers’ personality traits also significantly affect how they perceive difficulty in raising their child. When considering a caregiver of the 2-0 code type, for example, one may be challenged by their child’s aspects that make one feels tough to raise, such as aggressive behavior, because these people tend to represent chronic depression, guilty, social withdrawal, and lack of confidence (Levak et al., 2012). Additionally, higher scale scores of 2-0 code type are clinically associated with unipolar depression or remitted depression, and pessimism, the negative cognition of which might lead to a bias that affects measuring a child as a more difficult child (Wetzler et al., 1995; Suzuki et al., 2014).

Total parenting stress measures the overall stress level in the role of parents (Abidin, 1990). The main contributors to the overall stress of caregivers were the ASD patients’ behavioral problems (aggressive behavior, anxious/depressed, attention problems, withdrawn, and other problems) and caregivers’ psychopathology [depression (code 2), social introversion (code 0), psychasthenia (code 7), and schizophrenia (code 8)], whereas ASD core symptoms showed less predictive power. This suggested that individualized interventions for caregivers targeting their mental health in the context of their psychopathologic profile might be helpful in alleviating parenting stress and subsequently improving treatment outcomes for ASD patients (Osborne et al., 2008). Indeed, the current interventions of mainstream generally aimed at maximizing ASD patients’ functioning by improving their core symptoms or specific behavioral problems (Lai et al., 2014). Psychological interventions targeting parenting stress were also suggested, of which the efficacy was confirmed by a meta-analysis of 16 randomized controlled trials with moderate certainty of evidence (standardized mean difference −0.33, 95% CI −0.46 to −0.19) (Kulasinghe et al., 2022). However, none of the meta-analyzed trials have included caregivers’ psychopathology as a treatment target. Our findings that caregivers’ psychopathology was also a reliable predictor of total parenting stress suggested that interventions may benefit from including caregivers’ psychopathology as a novel therapeutic target.

Meanwhile, there is a tendency that in the CBCL 1.5–5 sample, model features associated with the caregivers’ psychopathology were ranked higher than behavioral problems; however, this was inverted in the CBCL 6–18 sample. This might result from the tendency that the severity of behavioral problems tended to be higher in the CBCL 6–18 sample than in CBCL 1.5–5. For subgroup analysis by comorbid ADHD, the model features related to ASD core symptoms rose in rank in the sample with ADHD compared to the main analysis. This may indicate that addictive deficits in the social domain of both ASD and ADHD have more contributed to parenting stress than the ousted predictors (Mikami et al., 2019).

This study has some limitations. First, the performance of ML models was not confirmed by the external validation set. Second, given that generated models did not show the perfect performance, failing to achieve an AUC larger than 0.90, it could be hypothesized that some potential predictors of parenting stress, such as family income, may have been missed. Third, the thorough interpretation of the MMPI was not possible because the MMPI should be addressed in the context of validity scales, but only clinical scales were employed as model features. Fourth, the SRS-2 and CBCL were rated by a caregiver, possibly leading to the overestimation of the ASD patients’ status, especially in those with high parenting stress (Schwartzman et al., 2021). Lastly, the study period was insufficient to investigate the impact of the COVID-19 pandemic on the parenting stress of this population, which calls for further studies.

In conclusion, we identified major predictors for each component of parenting stress in ASD patients’ primary caregivers using explainable ML models. This study revealed specific components of caregivers’ psychopathology, ASD patients’ core symptoms, and behavioral problems which mainly contribute to parenting stress. Our ML models and the identified predictors would be helpful in alerting physicians whether a caregiver is at a high risk of experiencing severe parenting stress and if so, providing timely interventions, which could eventually improve the treatment outcome for ASD patients.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by Institutional Review Board of the Severance Hospital of Yonsei University, Seoul, the Republic of Korea. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because informed consent was waived since the authors used retrospective and deidentified patient data.

HC, JK, HK, and K-AC conceived and designed the study. HC managed data collection and checked data eligibility. JK performed the machine learning-based analysis. HC and HK inspected the completion of the analysis. HC and JK drafted the manuscript. HK and K-AC provided critical revision of the drafted manuscript. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.