Exploring Primary and Interaction Effects of Minor Physical Anomalies: Development and Validation of Prediction Models Using Explainable Machine Learning Algorithms for Early-Onset Schizophrenia

Abstract

Background and Hypothesis

Minor physical abnormalities (MPAs) are neurodevelopmental markers that can be traced to prenatal events and may be significant features of early-onset schizophrenia (EOS). Therefore, our study aimed to (1) find the primary and interaction effects of MPAs for EOS and (2) develop and validate the model for EOS based on explainable machine learning algorithms.

Study Design

The study included 549 patients with schizophrenia (193 EOS and 356 AOS) and 420 healthy controls (HC) in southern Taiwan. For the feature selection, variable selection using random forests (varSelRF) and recursive feature elimination (RFE) were applied to identify the important variables of MPAs. We used different machine learning algorithms to build the prediction models based on the selected MPAs variables.

Study Results

The results showed that the mouth anomalies are significant MPAs variables and have interaction effects with craniofacial MPAs variables for EOS. The prediction models using the selected MPAs variables performed better in discriminating EOS vs HC compared to AOS vs HC. The AUC values for distinguishing EOS vs HC were 0.85-0.93, AOS vs HC were 0.80-0.87, and EOS vs AOS were 0.67-0.77 in validation sets.

Conclusions

This risk prediction model provides a clinical decision support system for detecting patients at high risk of developing EOS and enables early intervention in clinical practice.

Introduction

Schizophrenia is commonly viewed as the clinical outcome of abnormal neurodevelopment influenced by both genetic and environmental factors.¹ Age of onset has been considered a significant feature of patients with schizophrenia and provides important clues for genetic origins.² Early-onset schizophrenia (EOS) begins in childhood or adolescence. EOS was associated with more severe symptoms, poorer socio-occupational functioning, more premorbid abnormalities, greater genetic loading, and more developmental deviance.^3–5

Minor physical anomalies (MPAs) are the neurodevelopment markers which are slight defects of the head, eyes, ears, mouth, hands, and feet.⁶ These anomalies are abnormal markers with genetic origins that are associated with gestational developmental insults.⁷ These neurodevelopmental markers occur during the first and early second trimesters of gestation and persist into adult life.^8,9 In previous studies, patients with schizophrenia usually displayed higher frequencies of MPAs compared to healthy controls (HC).¹⁰ Some studies of the age of onset of schizophrenia showed that many variables of MPAs are significantly different between EOS and adult-onset schizophrenia (AOS).^11–14 Although these studies confirmed that patients with EOS are more likely to have MPAs, most of the studies had small sample sizes and only measured the qualitative variables of MPAs, or only used summary scores of the qualitative variables or total region scores of the MPAs variables. Therefore, further studies with larger samples, combining qualitative and quantitative variables, and more detailed variables instead of summary scores may help researchers in more accurately identifying important variables for EOS. Furthermore, previous MPAs studies on EOS were focused solely on developing models to differentiate EOS from HC. Studies have been relatively few in establishing models to differentiate EOS from AOS.

As prediction models become more popular in clinical decision-making, machine learning is often used to build these models, rather than traditional methods. In epidemiology, machine learning can help to process high-dimensional data and predict the disease.¹⁵ Compared with traditional statistical modeling, machine learning methods can handle more potential predictors and perform better.^16,17 Another advantage of machine learning is that some methods can avoid multicollinearity of independent variables such as random forest and elastic net regression.¹⁸ To date, only one study has used machine learning to build the prediction model to detect important MPAs variables for distinguish EOS from HC.¹⁴ However, previous studies did not verify the extrapolation of the model for EOS and did not analyze the interaction of MPAs variables.

In summary, MPAs are markers of abnormalities in early fetal development that are often found in patients with schizophrenia and appear to be consistent with genetic liabilities. EOS may have a higher genetic loading and exhibit a higher prevalence of MPAs than AOS. Our study aimed (1) to find the primary and interaction effects of MPAs for EOS and (2) to develop and validate the model for EOS based on explainable machine learning algorithms.

Methods

Participants

For our study, we collected 549 patients with schizophrenia and 420 HC from April 2011 to May 2024. We recruited patients with schizophrenia from 6 hospitals in southern Taiwan. We recruited schizophrenia patients from 6 hospitals in southern Taiwan. Three hospitals (Jianan Mental Hospital, Lok An Hospital, and National Cheng Kung University Hospital) were assigned to the training set, while the other three hospitals (Chi Mei Medical Center, An-Nan Hospital, and National Cheng Kung University Hospital, Dou-Liou Branch) were assigned to the validation set. This resulted in 311 schizophrenia patients in the training set and 238 in the validation set. To match the case-control ratio, 55% of the control group (n = 231) was randomly allocated to the training set and 45% (n = 189) to the validation set. Patients with schizophrenia were recruited from inpatient and outpatient departments; these patients met the criteria for schizophrenia in the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV-TR, 2011-2018) or Fifth Edition (DSM-5, 2018-2023). The HC were recruited from the hospital staff or individuals in the community and had no history of psychiatric disorders. A history of illegal substance or alcohol abuse, identifiable neurological disorders, intellectual disabilities, somatic disorders with neurological components, or a parent who was not Han Chinese were criteria for exclusion from our study. We collected baseline information of participants such as sex, age, weight, height, body mass index (BMI), age of onset, and duration of illness. All participants provided written informed consent to be included in this study, which was approved by the Institutional Review Board of the participating hospitals.

Age of Onset of Schizophrenia

In our study, the age of onset was defined as the age at which the patient’s symptoms met DSM-IV-TR or DSM-5 criteria. If patients with schizophrenia were diagnosed before the age of 20, they were classified as EOS. The cutoff points most frequently used for EOS are 18 and 20 years old.^2,19 We used a cutoff point of 18 or 20 years old to distinguish between EOS and AOS and found the results were similar (Supplementary Table S1). To ensure a larger sample size and more stable results, we used 20 years old as the cutoff point for EOS.

Measurement of MPAs

The scale used by our study combines qualitative and quantitative variables. The qualitative variables are based on previous studies.^10,20,21 The scale is divided into 6 regions: head, eyes, ears, mouth, hands, and feet, with 41 questions in total. Thirty-three variables evaluated the presence or absence of anomalies, and the other 8 variables used an ordinal scale (0, 1, and 2) to rate the degree of the anomalies. The quantitative measured MPAs scale contained 27 variables; this scale was developed by Lin et al.²⁰ These quantitative variables were separate left- and right-side measurements of eleven bilateral anatomical sites. We use 2 reference books on anthropometric measurements^22,23 to construct a standardized set of measurements. In our previous study, we evaluated the inter-rater reliability of twenty healthy subjects; the results showed that the inter-rater reliability for the qualitative variables ranged over 0.95-1.0, and the range of inter-class correlation coefficients was 0.70-0.99 for the quantitative measurement variables.¹⁴

Feature Selection

Variable Selection Using Random Forest

Variable selection using random forests (varSelRF) is a backward variable-selection method based on random forests. It calculates the important variables and removes the variables at each iteration,²⁴ calculates the out-of-bag error in each iteration and selects the variables when the corresponding RF model achieves the lowest misclassification rate.²⁵ This method removes the least important features while maintaining a similar error rate to the full model.²⁶

Recursive Feature Elimination

Recursive feature elimination (RFE) is a method that is typically used together with many classification algorithms.²⁷ The classifier estimator is trained using the initial set of features and these features are sorted according to their weights. Features with the smallest weights are removed.²⁸ In this study, we used the RFE method based on the random-forest model to select the important variables. Compared with other methods, random forest RFE (RF-RFE) has been shown to use fewer variables to achieve a higher classification accuracy.²⁹

Prediction Models

Logistic Regression

Logistic regression is a traditional analysis method for a binary event and it can deal with classification problems. The aim of logistic regression is to map a function from the features of the dataset to the targets in order to predict the probability that a new example belongs to one of the target classes.³⁰

Elastic Net Regression

Elastic net regression is a regularized regression, it combines lasso and ridge regressions to minimize the loss function. This method can introduce penalized covariates that do not significantly improve the fit of the model and are subsequently shrunk until they are forced out of the model entirely.³¹

Random Forest

Random forest is a machine learning method similar to a decision tree and it can deal with supervised classification and regression task problems.³² It combines many decisions to predict the outcome and incorporates feature selection and interactions in the learning process. For classification tasks, the most frequent prediction (mode) across the decision trees is used as the final prediction. Random forests are used to reduce variance and overfitting associated with decision trees.³³

Support Vector Machine

Support vector machine (SVM) is a powerful classification learning model³⁴ that attempts to find a good separating hyperplane in a high-dimensional feature space.³⁵ This method maximizes the margin between the closest data points of opposite classes, making it effective for both classification and regression tasks. It finds an optimal hyperplane to separate data into different classes and can handle complex decision boundaries.³⁶ The advantages of SVM are robustness, good generalization ability, and unique global optimum solutions.³⁷

eXtreme Gradient Boosting

eXtreme gradient boosting (XGBoost) is an optimized distributed gradient boosting library designed to be efficient, flexible, and portable.³⁸ This method through an iterative calculation method to combine classifier groups with low accuracy to make a high-accuracy classifier. The characteristics of XGBoost are accurate training results, fast running speed, and loose data requirements.³⁹

Statistical Analysis

All statistical analyses were performed using SAS 9.4 (SAS Institute, Cary, NC) and R 4.2.3 (R Foundation, Vienna, Austria). The flowchart of data analysis is presented in Figure 1.

Descriptive Data

In the demographic and clinical variables analysis, we used the t-test and chi-square test to compare the differences between the 2 groups. A P-value <.05 was considered to indicate statistical significance.

Model Performance

The performance of the final prediction models was evaluated on the training set and the validation set, using 4 metrics to evaluate the performance: Area under the curve (AUC), sensitivity, specificity, and accuracy. We utilized bootstrapping analysis to estimate the bootstrap AUC with 95% confidence interval by 1000 bootstrapping resampling, providing the evaluation of internal validity. Furthermore, we used 10-fold cross-validation to assess the predictive ability of the model and to determine whether the model exhibits overfitting. Additionally, we used the SHAP (SHapley Additive exPlanations) value to evaluate the important features of MPAs in the XGBoost model. Moreover, our study accounted for group-level random effects of hospitals by employing a mixed-effects logistic regression model and GPBoost, a machine learning algorithm inspired by XGBoost, to assess their influence on predictive performance.

Evaluation of the Interaction Effect

We utilized the R package “iml” to detect the interaction effects of MPAs variables. This package employs Random Forest and the H-statistic to assess and generate pairwise interaction terms.⁴⁰

Results

Descriptive Statistics

Supplementary Table S2 presented the demographic and clinical characteristics of the age of onset of schizophrenia in the training and validation sets. We found that sex was significantly different between EOS vs HC and AOS vs HC in the training set. Age, weight, and BMI were significantly different between EOS vs HC and AOS vs HC in both the training and validation sets. Age of onset and duration differed significantly between EOS vs AOS in both the training and validation sets.

Feature Selection of MPAs

We used 2 feature selection algorithms (varSelRF and RFE) to select the MPAs variables for EOS vs HC, AOS vs HC, and EOS vs AOS (Supplementary Figures S1–S3). The MPAs variables selected by the 2 feature selection algorithms simultaneously were used to construct the prediction model. Therefore, 14 variables, 13 variables, and 7 variables were considered to be the selected variables for EOS vs HC, AOS vs HC, and EOS vs AOS, respectively.

Performance of Models

The selected MPAs variables from the feature selection algorithms were used to evaluate the performance of the models. The discrimination ability of the risk prediction models based on logistic regression and 4 machine learning algorithms for EOS vs HC, AOS vs HC, and EOS vs AOS are illustrated in Table 1. The bootstrap AUC values of random forest, SVM, and XGBoost are shown in Figure 2. The AUC values of all models in EOS vs HC were greater than 0.85, and XGBoost exhibited the highest performance, with the AUC of 0.93 in the validation set. The MPAs variables performed better in discriminating EOS vs HC compared to AOS vs HC. The AUC values of EOS vs AOS using different machine learning algorithms were 0.67-0.77. The predictive ability for the EOS vs HC and AOS vs HC models, which included sex and BMI, were presented in Supplementary Table S3. We also demonstrate the model performance based on 10-fold cross-validation in Supplementary Table S4. The AUC values of EOS vs HC, AOS vs HC, and EOS vs AOS were 0.85-0.87, 0.79-0.82, and 0.66-0.68, respectively. Considering the group-level random effects of hospitals, the AUC values for mixed-effects logistic regression were similar to those for logistic regression. When utilizing a mixed-effects machine learning algorithm based on XGBoost, the AUC value of GPBoost was 0.78, which closely matched the AUC of XGBoost (0.77) (Supplementary Figure S4).

To evaluate the importance of MPAs variables in the model, we selected the model with the highest AUC for EOS vs HC, AOS vs HC, and EOS vs AOS. The importance of MPAs variables in XGBoost for EOS vs HC, AOS vs HC, and EOS vs AOS is illustrated in Figure 3. The X-axis indicates the SHAP value and the Y-axis indicates the importance of the MPAs variables in making predictions. Each dot represents one patient in the validation set. We found that, in EOS vs HC, the top 3 significant MPAs variables were facial width, high/steepled palate, and Hyperconvex fingernails. In AOS vs HC, the top 3 significant MPAs variables were epicanthus, high/steepled palate, and tongue with smooth–rough spots. In EOS vs AOS, length of the philtrum, tragion to subnasale, and facial width were the top 3 significant MPAs variables. We also calculated the effect sizes of the important MPA variables in EOS vs HC, AOS vs HC, and EOS vs AOS (Supplementary Figure S5). The results demonstrated that most of the important MPA variables showed significant differences between the 2 groups in EOS vs HC, AOS vs HC, and EOS vs AOS.

Interaction Effects of MPAs Variables

The interaction effects of EOS vs HC and AOS vs HC are shown in Supplementary Figures S6 and S7, respectively, while the interaction effect of EOS vs AOS is displayed in Figure 4. Mouth anomalies and craniofacial anomalies were the MPAs variables that easily interacted with other MPAs variables in both EOS vs HC and AOS vs HC comparisons. The length of the philtrum was identified as an MPAs variable likely to interact with other MPAs variables in EOS vs AOS. Interpupillary distance and tragion to subnasale were also identified as MPAs variables likely to have interaction effects with other selected MPAs variables in EOS vs AOS. These findings suggest that mouth anomalies and craniofacial anomalies may interact with other MPAs variables. Furthermore, to better understand the interaction effect, we selected 2 MPAs variables from EOS vs HC, AOS vs HC, and EOS vs AOS as examples to present the bivariate distribution of 2 MPAs variables with interaction in Supplementary Figure S8.

Discussion

To the best of our knowledge, our study is the first to use machine learning algorithms to identify the primary and interaction effects of MPAs with the age of onset for schizophrenia and develop the risk prediction model for EOS vs AOS. We combined qualitative and quantitative MPAs variables to select the important variables by applying 2 feature selection algorithms. We also used the selected MPAs variables to develop and validate the risk prediction model for EOS. The mouth anomalies are the MPAs variables that are important for discriminating between schizophrenia patients and HC and easily have interaction effects with craniofacial MPAs when comparing EOS and HC. Our study compared different machine learning algorithms to develop the risk prediction for EOS and showed high discriminative abilities in the validation set. The risk prediction model can assist in identifying individuals at high risk of schizophrenia for clinical practice or research purposes.

MPAs are an indirect index of early prenatal central nervous system maldevelopment and serve as markers of altered morphogenesis occurring during the first or early second trimester of pregnancy.^41–43 A prospective cohort study examining the link between childhood MPAs and adult psychiatric outcomes found that children with a high number of MPAs had an increased likelihood of developing schizophrenia spectrum disorders compared to other psychiatric conditions.⁴⁴ Previous studies revealed that EOS exhibited more MPAs than AOS, suggesting a stronger association between MPAs and EOS.^11,12,14 This association emphasizes the potential relationship between early prenatal central nervous system maldevelopment and the onset of EOS. Our results also demonstrated that the predictive ability of the EOS model was better than the AOS model. On the other hand, the great predictive performance of EOS vs AOS also suggested that EOS had more MPAs than AOS. Namely, MPAs may serve as markers of ectodermal maldevelopment during embryonic gestation, associated with genetics and acting as predictive indicators for EOS. EOS is characterized by more severe symptoms, diminished socio-occupational functioning, notable premorbid abnormalities, higher genetic susceptibility, and greater developmental deviations.^3,4,45 Developing a risk prediction model based on MPAs for EOS holds significant clinical value by improving disease prediction and enabling early interventions. This predictive model supports the early diagnosis of schizophrenia, facilitating the delivery of more personalized and effective treatment strategies. Furthermore, it aids in identifying high-risk groups for EOS, providing a framework for implementing precise and proactive prevention efforts.

In our study, we identified that a furrowed tongue and high/steepled palate can discriminate schizophrenia vs HC. Previous studies suggested that mouth anomalies and a high/steepled palate were important variables with good ability to discriminate between EOS and HC.^12,14 Studies have indicated that palate anomalies are more commonly observed in schizophrenia patients, and the prevalence of high palate anomalies varies across different types of schizophrenia.^46,47 A study by Tsai et al. found that hyperconvex fingernails and facial width were associated with EOS, which is consistent with our findings. Recent studies found that MPAs related to schizophrenia are concentrated in the mouth, head, and other craniofacial regions.^7,48 Our study found similar results, which may be explained through the neurodevelopmental hypothesis of schizophrenia. Craniofacial features are related to underlying brain development due to shared ectodermal origins; therefore, craniofacial features may be useful biomarkers for predicting schizophrenia onset.^49–52 In addition, according to a previous study, the development of midline craniofacial structures occurs between gestational weeks 9-10 to 14-15. The findings explain the neurodevelopmental basis of schizophrenia and strongly support the notion that risk factors operate during organogenesis in the first trimester of pregnancy.⁵³

For the interaction effect in EOS, previous studies reported that craniofacial anomalies may be affected by each other during embryonic development.^54,55 Taking an embryological perspective, the cranium and face collectively provide support to the developing brain, and disruptions in the typical growth of cranial structures are intricately linked to disruptions in the normal development of the brain.^56,57 Recent studies of facial shape and craniofacial structure indicated that high palate, facial width, palpebral fissure length, and length of mouth were heritable facial features.^58,59 Our results were consistent with previous studies showing that mouth variables were likely to interact with other craniofacial MPAs variables.

Recent meta-analyses have demonstrated that machine learning algorithms can achieve significantly better discrimination ability than logistic regression in predicting diseases.^60–63 Among the models in our study, random forest, SVM, and XGBoost demonstrated better prediction performance. Interpreting the prediction model of machine learning correctly and presenting the results visually is a constant challenge. Therefore, we used the SHAP value for model interpretability. The SHAP value can provide explanations not only at the global level for the entire dataset but also at the individual level, offering insights into which variables influence specific prediction results for each patient and helping users understand the reasons behind the model’s predictions.^64,65

The limitations of this study are as follows. First, this study included only Taiwanese participants, and whether these findings apply to other ethnic groups remains to be explored. Second, qualitative MPAs variables usually do not change over time, but quantitative MPAs variables may require further confirmation. Third, we used linear measurements for quantitative MPAs variables, which may not provide a complete assessment of 3-dimensional structures, for example, the craniofacial region. Fourth, although MPAs are biomarkers of deviation in morphological development, MPAs measurement may not directly detect underlying neurodevelopmental disturbances.

In conclusion, our study used machine learning algorithms to develop and validate the model for EOS. Important MPAs variables were selected by feature selection methods based on machine learning algorithms and used to build a prediction model for EOS vs HC, AOS vs HC, and EOS vs AOS. Machine learning algorithms exhibited a higher model performance in EOS vs HC than in AOS vs HC and demonstrated effective performance in differentiating EOS from AOS. Mouth anomalies are significant MPAs variables and easily interact with other craniofacial MPAs, whether in distinguishing EOS from HC, AOS from HC, or EOS from AOS. These findings assisted us in identifying important variables in MPAs and evaluating the interaction effects that enhance the risk associated with the onset of schizophrenia. This risk prediction model shows promise as an effective clinical decision-making tool, enabling the early identification of patients at high risk for schizophrenia and supporting timely interventions in clinical practice. Moreover, high-performance predictive models promote personalized treatment strategies, minimizing unnecessary pharmacological interventions. Through this model, we expect to assist clinicians in diagnosing schizophrenia more accurately and arranging effective, individualized treatments for each patient.

Acknowledgments

The authors sincerely appreciate the assistance of all laboratory members, especially research assistants Ya-Hsin Liu and Chien-Chien Chen.

Author Contributions

C.-W.L.: Writing—Original Draft, Visualization, Software, Methodology, Formal analysis, Data Curation, Validation, Conceptualization. J.-J.L.: Resources, Methodology, Investigation. H.-H.T.: Resources, Methodology, Investigation. F.-L.J.: Resources, Methodology, Investigation. M.-K.L.: Resources, Methodology, Investigation. P.-S.C.: Resources, Methodology, Investigation. C.-C.H.: Resources, Methodology, Investigation. C.-Y.Y.: Resources, Methodology, Investigation. T.-Y.W.: Resources, Methodology, Investigation. W.-H.C.: Resources, Methodology, Investigation. H.-P.T.: Resources, Methodology, Investigation. S.-H.L.: Writing—review & editing, Supervision, Resources, Methodology, Funding acquisition, Data curation, Validation, Conceptualization.

Funding

Funding for this study was provided by the National Science and Technology Council of Taiwan (NSTC 113-2314-B-006-072-MY3, NSTC 113-2321-B-006-014, and NSTC 112-2314-B-006-024) and the Ministry of Science and Technology of Taiwan (MOST 107-2314-B-006-066, MOST 108-2628-B-006-015, and MOST 109-2314-B-006-054-MY3).

Conflicts of Interest

None declared.

References