A multimodal prediction model for suicidal attempter in major depressive disorder

Study populations

The participants used in this study were from the UK biobank (UKBB) (http://www.ukbiobank.ac.uk), which is a population-based cohort including over 500,000 participants recruited in the United Kingdom between 2006 and 2010 (Sudlow et al., 2015). Among the 502,616 participants, exclusively adults, with age ranges from 40 to 77, the mean age at baseline was 59.46 years (standard deviation 8.12), 54.41% were men and 81.51% were of White ethnicity. UKBB is used to improve the prevention, diagnosis, and treatment of various diseases. The baseline assessment includes lifestyle, environment, medical history, genomics, physical measures, and other relevant data. Informed consent was obtained from all UKBB participants.

This study was subject to UKBB ethical approval granted by the National Information Governance Board for Health and Social Care and the NHS North West Multicenter Research Ethics Committee. All participants gave informed consent by electronic signature at baseline. Data collected at baseline were used in this study. This study was conducted using the UK Biobank Resource under Application Number 75556.

Definition of major depressive disorders

The definition of MDD was based on major depression status in the UKBB(field ID: 20126). The definition of MDD in this field is generated based on the hypothetical categories of MDD summarized by Smith et al. (2013), which are classified into single-episode major depression, moderate recurrent major depression, and severe recurrent major depression. For participants with missing data, we excluded them from the study. Finally, there were a total of 31,829 MDD patients included in this study.

Definition of suicidal attempts

The definition of suicidal attempts was based on “Ever attempted suicide” (field ID: 20483). Each participant was asked to answer a question about “Have you harmed yourself with the intention to end your life?”. In the latest data released by UKBB, excluding those participants with the answer “Prefer not to answer” ( n = 219), there were a total of 6,642 participants answering this question. Finally, there were a total of 1132 MDD patients with completed the suicidal attempts questionnaire. Among 1,132 MDD patients, there were a total of 653 participants defined as MDD patients with suicide attempts (MDD-SA, n = 653), and a total of 479 participants defined as MDD patients without suicide attempts (MDD-nSA, n = 479).

Features characteristics

A total of 179 features were included in the prediction analysis, including demographic features (n = 11), depressive symptoms features (n = 15), and brain structural phenotypes features( n = 153). More information about these features is described in Table S1.

Demographic features. There were eleven demographic features, including gender, age at the time of attending the assessment center, current employment status, age at completion of full-time education, moral background, age at the time of the first episode of depression, age at the time of the last episode of depression, sleep duration, smoking status, variation in diet and alcohol intake frequency.

Depressive symptoms feature. There were fifteen depressive symptoms features, for example, “numbers of depression episodes”, “even had prolonged feelings of sadness or depression”, “feeling of tiredness during a worst episode of depression”, and so on. More information about these features is described in Table S1.

Brain structural phenotypes. In this study, neuroimaging data were scanned at the UK Biobank Imaging Centre by a standard Siemens Skyra 3T scanner running VD13A SP4 and a standard Siemens 32-channel radiofrequency receiver head coil (Miller et al., 2016). More information is shown at https://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf. The FAST and FIRST grey matter segmentation are used to generate a further 153 regional brain structural phenotypes, by summing the grey matter partial volume estimates within 153 ROIs: 139 regional grey matter volumes (GMV) and 14 subcortical volumes (Category ID: 1101 and 1102). These ROIs are established in MNI152 space by merging several different atlases: the Harvard-Oxford cortical and subcortical atlases (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Atlases) and the Diedrichsen cerebellar atlas (http://www.diedrichsenlab.org/imaging/propatlas.htm). More details are available at https://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf. A total of 42,789 participants’ neuroimaging data were used for the present study. Due to the small amount of data, we excluded fields with missing rates exceeding 10% from the study. For the remaining categorical variables, the missing values were filled in using the mode value. For the remaining continuous variables, the missing values were filled in using mean and then standardized to z-score for further analyses.

Finally, 208 MDD patients (MDD-SA, n = 119; MDD-nSA, n = 89) with complete 11 demographic features, 15 depressive symptoms features, and 153 brain structural phenotypes features were included in the further analysis. A flowchart for describing the sample size in each analysis is shown in Fig. S1.

Statistical analysis

In this study, an ML multimodal model was built for the prediction of suicide attempters in MDD patients by integrating SVM-RFE and RF algorithms in 208 MDD patients. The process includes feature selection, feature integration, and prediction model estimation. The working flowchart in this study is described in Fig. 1 and details are introduced below.

Feature selection

Feature selection is an important step in the classification prediction model. In this study, two classical feature selection methods, namely SVM-RFE and RF, were used to explore important features in the classification of MDD-SA and MDD-nSA. This study implemented ML feature selection using the scikit-learn package with Python 3.7 (http://scikit-learn.org/). SVM-RFE algorithm is a feature selection algorithm for recursive feature elimination based on the principle of maximum interval of SVM, in which the three most important parameters are kernel function, kernel function coefficient gamma, and penalty coefficient C. In this study, we set the kernel function to be “linear kernel”, the kernel function coefficient to be “auto”, and the penalty coefficient to be 1. RF is an algorithm for feature selection by calculating the contribution of each feature on each tree in a random forest, the two most important parameters are the number of decision trees and the classification criteria of the tree nodes criterion, which we set to 50000 and “Gini”, respectively. More information about the algorithm parameters used in the experiments is shown in Table S2.

The specific implementation process of the SVM-RFE method was as follows: (1) All 179 features were introduced into SVM-RFE to train the classifier, in which the less important features were iteratively eliminated according to the weighted vector of SVM with linear kernel, and the features were rearranged in descending order according to their importance; (2) Repeated training and testing split design was conducted using leave-one-out cross-validation (LOOCV). Specifically, in each experiment, one dataset was used for model test, and all other data was used for model training. The feature importance ranking was recorded according to all the iterative results. (3) Finally, the predictive features that survived from Subset-SVM-RFE were adopted for the following analysis.

The implementation of the RF method was as follows: (1) All 179 features were introduced into an RF classifier to build 50,000 subtrees, using Gini Impurity as an evaluation criterion for dividing the subtrees (Menze et al., 2009); (2) The features were ranked by importance and the top features were selected as subset-RF for subsequent analysis. The numbers of top features in Subset-RF were consistent with the numbers of features in Subset-SVM-RFE.

Feature integration

The features in Subset-SVM-RFE and Subset-RF were then integrated for further analysis. The common features in Subset-SVM-RFE and Subset-RF were treated as Features-Subset and other features were added to the Features-Subset one by one based on their importance for model training. The dataset for model training was stratified into train and test datasets with a ratio of 7:3, ensuring the category proportion. A linear kernel was used to fit the SVM model, the rest of the parameters was based on the default parameters of the SVM model. More information about the algorithm parameters used in the experiments is shown in Table S2.

Prediction model estimation

Three indices in prediction model estimation, including accuracy, sensitivity, and specificity, were used to estimate the prediction model. Additionally, two important statistics, positive predictive value (PPV) and negative predictive value (NPV) were used to measure the optimized clinical relevance of a test. The receiver operating characteristic (ROC) curve for this optimal model was also demonstrated. To validate the selected features that survived from the integrated prediction model, a two-sample T-test of the predictors in the resulting optimal classifier was performed to compare the differences in features among MDD-SA and MDD-nSA participants.

Demographic

A total of 208 MDD patients were finally included in this study, their demographic and clinical characteristics are shown in Table 1. There were no statistical differences between MDD-SA (n = 119) and MDD-nSA (n = 89) in terms of sex, current employment status, sleep duration, smoking status, dietary changes, and frequency of alcohol consumption (Table 1). The age at the time of assessment in the MDD-SA group (53.09 ± 7.51 years) was significantly higher than those in the MDD-nSA group (50.54 ± 7.07 years) (P = 0.015).

Feature selection and integration

In the feature selection using SVM-RFE, the Subset-SVM-RFE selected a total of 21 predictive features. Correspondingly, the Subset-RF included the top 21 predictive features from the RF algorithm (Table S3). The SVM-RFE and RF algorithms were integrated to explore useful prediction features for the classification of MDD-SA and MDD-nSA. We found that there were eight common features between Subset-SVM-RFE and Subset-RF. The other 34 features from Subset-SVM-RFE and Subset-RF were added to the predictive model interactively to estimate the predictive performance. More information about these features is shown in Table S3.

Predictive model performance

In this study, three model prediction indices, including accuracy, sensitivity, and specificity, were used to estimate the predictive model. The process of adding features one by one to achieve the best classification effect is shown in Fig. 2. As shown in Table 2, in the prediction model of a single feature selection method, RF demonstrated an accuracy of 65.08% of predictive performance, and SVM-RFE demonstrated an accuracy of 76.19%. After iteratively integrating features from Sub-SVM-RFE and Sub-RF, we found that the predictive model with 34 features demonstrated the best performance, with a balanced accuracy of 77.78%, a sensitivity of 83.33%, and a specificity of 70.37%. Additionally, two important statistics, PPV and NPV were used to measure the best clinical relevance of a test. In this study, the PPV of MDD-SA was 78.95%, and the NPV value was 76.00%. The ROC curve with AUC =0.75 for this optimal model is shown in Fig. 2. For the prediction model of integrated feature selection methods, its prediction accuracy is nearly 13% higher than that of feature selection using RF, and 1.59% higher than that of the model using SVM-RFE for feature selection. This may be because the SVM-RFE algorithm calculates feature scores based on SVM, and the selected features can fit the model well when reintroduced into the SVM predictor for prediction, resulting in roughly equal accuracy between the two algorithms.

This study additionally conducted a post-hoc analysis on the finally used features in the model, and significant differences were found in 9 features from the above 34 features between MDD-SA and MDD-nSA. The brain structural phenotypes feature included bilatary precuneus cortex volume, bilateral hippocampal volume, left supracalcarine cortex volume, and right thalamus volume. The demographic and behavioral features included “age when attending assessment centre”, “number of depression episodes”, and “feelings of worthlessness during worst period of depression” (Table 3).