1 INTRODUCTION

Schizophrenia and bipolar disorder (BD) are critical and high-risk inherited mental disorders that have debilitating symptoms.¹ These disorders are among the severe psychiatric diseases that have many overlaps and similarities with each other and affect the patient's behavior in the family and society. According to the World Health Organization, these disorders are among the top 10 causes of disability worldwide.² Schizophrenia and BD affect 3% of the world's population.^{3, 4} Patients with schizophrenia and BD have a higher mortality rate than the general population.⁵ One of the prominent causes of death in these patients is suicide. In Danish registers, the rate of suicide is reported as 7.8% in men and 4.9% in women with BD.⁶

There are Five key features that define schizophrenia and BD and other psychotic disorders. These include delusions, hallucinations, disorganized thinking (inferred from speech), grossly disorganized or abnormal motor behavior, and negative symptoms.⁷ As compared with other disorders, schizophrenia and BD specifically, the presence or absence of specific psychotic symptoms identified as first-rank symptoms (auditory hallucinations; thought withdrawal, insertion, or interruption; thought broadcasting; somatic hallucinations; delusional perception; feelings or actions controlled by external agents) may be particularly helpful for making the diagnosis.^8-10

Many patients with schizophrenia and BD experience a long clinical period.⁷ Symptoms of the disease begin between the ages of 16 and 30. These symptoms fall into three categories: positive (hallucinations, delusions, and mental disorders), negative (lack or absence of facial expressions, feelings of little pleasure, and decreased sense of speech), and cognitive (concentrating and maintaining difficulty).^8-10 BD patients experience persistent changes in brain structures, such as enlargement of the third and lateral ventricles of the brain and a decrease in the volume of gray matter in the anterior and middle cerebral cortex, cortical and mesotemporal cortex, and decrease in the posterior abdominal callus volume.¹¹ The economic costs associated with the disease vary from $94 million to $102 billion each year.¹² Therefore, this disease imposes a heavy financial burden on the patients, their families, and society.¹³ Predicting the disease can go a long way in preventing it and controlling its costs. Since it takes an average of 10 years from the onset of the first symptoms to the definitive diagnosis of the disease,¹⁴ current approaches cannot effectively diagnose these diseases. Machine learning (ML) techniques are proposed as an effective tool to meet this need.

ML is a domain of artificial intelligence that allows computer algorithms to learn patterns by studying data directly without being explicitly programmed.¹⁵ Artificial intelligence using ML is entering the realm of medicine at an increasing pace and has been tested in various clinical applications ranging from diagnosis to outcome prediction.¹⁶ The utilization of ML techniques has many advantages, such as recognizing diseases, reducing physician decision-making errors, reducing healthcare costs, and improving the performance of healthcare providers.¹⁷

Various models have contributed significantly to the health domain, from rule-based systems to advanced ML models (deep learning). These models have been used in prediction, diagnosis, and treatment in healthcare, such as predicting survival in breast cancer,¹⁸ diagnosis and prognosis of COVID-19,^{19, 20} level of lung cancer,²¹ etc. ML techniques are also used to diagnose, classify, and predict schizophrenia and BD.^22-26 Several ML methods have been used to predict the negative symptoms of schizophrenia based on speech signals^{22, 27} and to predict the recurrence of schizophrenia.²⁴ Also, many studies have been conducted on the extraction of various features of computed tomography scans and magnetic resonance imaging (MRI) images in the diagnosis and prevention of schizophrenia and BD.^28-30

Several algorithms, such as random forests (RF), support vector machine (SVM), and gradient boosting (GB), have been frequently used in this area. The RF method is fast, adaptable, and reliable for mining high-dimensional data. As the name suggests, an ensemble of many decision trees makes up a RF. The RF produces a classification for each tree, and the class voted on the most becomes the prediction.³¹ SVMs are linear models for classification and regression problems. Several practical problems can be solved through this technique, including linear and nonlinear problems. This algorithm generates a line or a hyperplane to classify the data into classes.³² A GB algorithm trains many models (typically decision trees) in sequential and additive order. The purpose of boosting is to transform weak classifiers into strong classifiers. Each new model in GB is designed to minimize prediction error as much as possible.³³

This study aimed to conduct a systematic review of ML algorithms for predicting schizophrenia and BD to help early and timely diagnosis of the diseases to improve patients’ health.

2 MATERIALS AND METHODS

3 RESULTS

We retrieved 1243 papers through database searches. After title and abstract screening 144 papers were identified for full-text assessment. Full-text assessment excluded 129 studies due to various reasons. Fifteen papers met the inclusion criteria (Figure 1). An analysis of the algorithms applied, the inputs they were trained on, the outputs they were trained to predict, and their relative performance statistics are presented. An average number of 185 patients were used in each study (mean = 681.4, SD = 1289.65). The median number of ML algorithms employed in each study was two (mean = 2.67, SD = 1.99). Most of the studies (n = 13) applied ML algorithms only to schizophrenia disorder, and one study applied ML algorithms to both schizophrenia and BDs. All the included studies applied ML algorithms to predict the symptoms of these disorders. Data utilizes in prediction of 10 studies was based on MRI. In eight studies, ML algorithms predicted schizophrenia disorder while in three studies these algorithms predicted symptoms severity.

4 DISCUSSION

In this study, the applications of ML to support clinical decision-making in schizophrenia and BD were reviewed. Results suggested that the use of ML in schizophrenia and BD is rapidly growing.^{51, 52} There is substantial room for further applications of ML technologies to schizophrenia and BD data. Many ML applications and modeling methods have been used based on the findings. A wide variety of successfully predicted outcomes could facilitate decision-making. The number of ML publications on schizophrenia and BD has rapidly increased over the past few years.^53-55

According to this study, most schizophrenia and BD ML publications were in the domain of MRI and fMRI data. Few ML studies have focused on EEG data. The applied modeling approaches were substantially heterogeneous. Variation in the studies was related to algorithms applied, input and output variables used, and methods for assessing predictive model performance.

In this paper, SVM and RF were the most commonly applied algorithms, compared to other studies, applying SVMS^56-59 and RF.^{60, 61} Consistent with a previous study,⁶² RF frequently outperformed most other algorithms on supervised learning tasks. Since both SVM and RF are discriminative, they can handle large amounts of data, and capture nonlinear relationships across input features.⁴⁴ They were selected to predict disease outcomes, often demographics, clinical history, and investigation-related features were used.⁴⁹ AUC, sensitivity, accuracy, specificity, and precision performance metrics were the most commonly reported.⁴⁹ The accuracy of RF was significantly higher than that of SVM, and compared to SVM, the precision of GB was significantly higher. To facilitate future modeling, the input and output variables were synthesized. Gaps in knowledge and opportunities concerning the development of additional clinical decision support systems to improve the care of schizophrenia and BD patients were highlighted to clinicians and data scientists. To our knowledge, this systematic review is the first attempt to characterize the deployment of ML methods in schizophrenia and BD. We identified two different topic clusters, and keywords for each of the schizophrenia and BD subdomains by applying ML technologies. The use of these topics and keywords in the systematic review helped to deeply understand the past foci of the relevant literature and determine the current state of the knowledge. This understanding helps to identify gaps to be filled out by future researches. ML provides a series of technologies, from linear regression to RF, that can effectively predict outcomes for improving clinical decision-making in schizophrenia and BD. Based on the results, wherever large, labeled datasets were available, RF has been the top-performing algorithm in the ML methods due to its higher accuracy than multivariate logistic regression.⁶²

RF is faster to train and can readily handle larger numbers of predictors. It employs fewer parameters⁷¹ and does not need cross-validation (it generates an internal unbiased estimate of the generalization error (test error) as the forest building progresses).

5 LIMITATIONS AND FUTURE RESEARCH

It is more likely to publish positive and significant findings.^{72, 73} Many of the included studies evaluated the performance of multiple algorithms and reported the inferior performance of some. However, since we could not find studies reporting failed ML modeling activities, publication bias and selective outcome reporting may have influenced our results. Few studies have comprehensively reported predictive model performance statistics. Hence, we had relatively small sample sizes for our statistical analysis of algorithm performance. Larger sample sizes could have demonstrated additional significant performance differences between algorithms. More research is suggested to examine AUC, accuracy, sensitivity, and specificity to facilitate robust model performance evaluation and compare studies. Future studies can contribute to the convergence of ML with schizophrenia and BDs. Future studies can use similar predictors and outcomes on bigger databases containing the information of patients with schizophrenia and BDs. This may help predict the same or additional outcomes that can be used to deploy systems based on these models. The number of software to predict and support schizophrenia and BDs decision-making is low. Further models have required these two disorders. Despite the richness of electronic health record datasets, they are not utilized sufficiently and are suitable for conducting ML studies. SVMs and transfer learning can utilize video data during real-time patient examination, diagnosis and prognosis, or intraoperative anatomical identification. SVMs can predict the postoperative status of patients on an hourly basis to provide safer proactive management of patient status.

6 CONCLUSION

ML effectively supports making clinical decisions before, during, and after the onset of schizophrenia and BDs. Due to the evident heterogeneity of apply ML in psychology and the popularity of distinctive ML studies, there is a substantial potential for future similar studies. Because of the accurate prediction of various operative outcomes, RF seems more effective than other algorithms in schizophrenia and BDs. Deploying ML by designing sound clinical decision support systems can lower complications and increase the quality and safety of health services provided to schizophrenia and BD patients.

AUTHOR CONTRIBUTIONS

Mahdieh Montazeri: project administration. Mitra Montazeri: formal analysis; writing—review & editing. Kambiz Bahaadinbeigy: conceptualization. Mohadeseh Montazeri: data curation; writing—original draft. Ali Afraz: conceptualization; writing—original draft.

CONFLICTS OF INTEREST

The authors declare no conflicts of interest.

TRANSPARENCY STATEMENT

The lead author Mitra Montazeri affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.