Bipolar disorder (BD) is a chronically progressive mental disorder with a prevalence that ranges from 1.1 to 2.4% (1, 2). BD is classified as type I if the patient has presented at least one manic episode, with or without depressive episodes, and as type II in the presence of at least one hypomanic episode, with no full manic episodes, and one major depressive episode (3). This condition is associated with a reduced quality of life and greater disability. Patients have been shown to have lower incomes, higher financial burdens, issues with social interactions and a greater overall frequency of use of health services (4–7). Furthermore, significant suicide attempt rates have been reported in patients with BD type I (36.3%) and BD type II (32.4%) (8, 9). Additionally, manic episodes give rise to destructive and reckless behavior secondary to an unstable mood, which in the long term can degrade cognitive functioning, interpersonal relationships, global functioning and social adjustment (10, 11).

Patient admissions are preventable events with a considerable impact on healthcare costs and a key quality metric for health systems around the world (12–14). Among psychiatric patients, whose disorders are characterized by chronicity and high recurrence rates, readmissions are of particular concern. The period immediately after a hospitalization is known to be a period of high risk for outcomes such as suicide and substance abuse relapse. Patient admission and readmission due to BD places a significant financial burden on medical services and caregivers (2, 15). Studies have reported relapse rates of up to 50% after 2 years in individuals receiving adequate psychopharmacological care (16). Self-monitoring data from the Sanitas Healthcare Management Organization (HMO) in Colombia show that in 2017, 319 patients with a BD diagnosis had a total of 427 hospital admissions (27% of the overall number of readmissions) with a mean hospital stay duration of 19 days.

The prediction of risk for patient admission and readmission may aid in disease management as well as in reducing the economic and social burdens caused by BD. Despite the considerable number of publications in the field of patient admission prediction, most studies address this issue in patients with non-psychiatric illness (17–21), substance abuse disorders, schizophrenia or postpartum depression (22–24). Prior work by Rotenberg et al. aimed to predict depressive relapses in patients with bipolar disorder using machine learning techniques, achieving F measures as high as 0.993 for a random forest model (25). Although certainly related, depressive relapses are only one potential form of relapse of BD and may indeed be less likely to require hospitalization than manic episodes. This is why this study developed prediction models for hospital admission/readmission within 5 years of diagnosis in patients with BD using ML techniques. With clinical data that is readily available in electronic health records, we show that random forest models are good predictors of both admission and time to admission in patients with BD.

The prediction of patient admissions in any chronic disease is of great importance to researchers, public health planners and administrators (38, 39). Due to the varying features of different health systems, it is necessary to obtain different prediction tools for each population and health system to provide adequate risk management. To the best of our knowledge, this is the first study aiming to predict this outcome in a sample of subjects with BD in Colombia. We studied the outcomes of admission and readmission in subjects with BD in a large sample of 2,726 patients in Colombia, obtaining models with excellent predictive performance for the outcome of admission.

The ML models used in this study outperformed traditional statistical techniques (LR and P-Cox) in all cases with the random forest being the superior model in all cases. However, when comparing the results for the outcome of admission the random forest is only marginally superior to the LR model (AUC 0.98 vs. 0.94, respectively), which has the advantage of the latter of providing an easily interpretable model. This interpretability may be of critical importance in decision making. The difference between traditional statistical models and ML became more evident in the survival models for admission, in which the P-Cox model achieved a C-index of 0.897, while the SRF reached a C-index of 0.95. This difference between traditional statistical models and ML models was still evident in the survival models for readmission, though both models had very poor performance. This poor performance is likely due to the intrinsic difficulties associated with the prediction of survival (no readmission) compared with merely predicting the occurrence of the event within a given time frame. In addition, holdout validation was used to split the dataset corresponding to this outcome, which contained a smaller number of patients, leading to less precise estimates.

Random forest models are almost always superior to individual decision trees due to a variety of reasons. First, the random forest combines multiple decision trees which allows for a reduced variance and over-adjustment inherent to isolated trees. More robust and generalizable models can be derived from the averaging of several decision trees. Furthermore, the random forest models can also capture non-linear relationships and interactions between features. Lastly, random forest models can efficiently handle large datasets with high dimensionality, which makes them an adequate solution to complex problems like the prediction of BD admissions and readmissions.

All of the models explored for the outcome of admission had good to excellent performance, which was in contrast to the readmission models. ML models are known to require larger datasets than traditional statistical methods in order to produce better relative performance, and despite over and undersampling, the training dataset for readmission was considerably smaller (40). Prospective validation for these models is also lacking and it is an area for further research. The results suggested that ML analytics has the potential to provide risk calculators to aid in predicting clinical prognosis (including patient admissions), for individual patients (41).

Two prior studies have aimed to predict admission or readmission in BD patients. The study by Salem et al. aimed to predict readmission within 30 days after inpatient treatment of patients with a Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV diagnosis of BD using a SVM technique. Importantly, this study relied on features extracted from the Borderline Personality Questionnaire (BPQ), the scores of which were available for all subjects. The study found good discriminative ability with an area under the Receiver Operating Curve (ROC) of 0.86, concluding that borderline personality features, as measured using the BPQ, were good predictors of early readmission. The external validity of this study is limited by the availability of data on its primary predictor, which is not a standard feature of electronic health records of patients with BD (42). A second study, by Edgcomb et al., also aimed to predict risk of readmission after 30 days in patients with BD. This study took into consideration standard data from EHRs, not including any kind of standardized measurement. Using classification trees, this model achieved high accuracy with an area under the ROC curve of 0.88. An additional strength of this study was the interpretability of the model which can be easily adapted into medical thinking (43).

In this study, the presence of borderline personality disorder features was significantly associated with admission in our first analyses. Prior work has identified these features as being a key predictor in rapid readmissions in patients with BD (42). Although this variable was not significantly associated with readmission, it did meet the requirements for inclusion as a feature in the predictive models. Similarly, some medical comorbidities were both significantly associated with admission (hypertension, hypothyroidism), and met criteria for inclusion in the readmission models. This highlights their importance in considering the risk of admission/readmission in patients with BD.

A number of clinical variables related to the patient’s mood, sleep quality, self-reported energy levels and medication adherence were not available to us, and they are known to be good predictors of both depressive and manic episodes in BD (43, 44). This was probably due to the paucity of information registered in EHRs which is more likely to contain specific changes in the disease’s symptoms and signs throughout clinical follow-up, instead of the evolution of all the clinical variables considered.

In obtaining the final sample, we were significantly limited by the availability of information. Despite having access to a larger cohort of patients with BD, data availability in EHRs limited their inclusion in the study. Both the number of registries and the information contained within them may have varied systematically. Physicians may be less inclined to describe clinical features in detail for patients who they deem low risk. Patients at greater risk of admission may also be those with poor adherence to outpatient follow-up, leading to fewer records from which to draw information. This may have led to selection bias in our sample.

Future work will concentrate on the prospective validation of the models obtained in this project, including the use of cross-validation methods (e.g., K-fold, leave one out) to allow more robust estimates of model performance by evaluating the model on different combinations of data. The limitations caused by the scarcity of data could be mitigated by EHR systems which can more intuitively allow the psychiatrist to rapidly register certain features of the mental exam, enabling more comprehensive work in mental health to be carried out. Additionally, the incorporation of longitudinal data, genetic analysis and dynamic modeling could facilitate the development of personalized treatment strategies that account for individual variations in disease progression and response to interventions in BD.

This study highlights the potential of utilizing ML techniques to predict hospital admission and readmission in patients with BD. The results demonstrate that ML models, particularly the Random Forest algorithm, exhibit superior predictive performance compared to traditional statistical methods. By leveraging EHRs and incorporating a range of sociodemographic and clinical variables, these models provide valuable insights into the factors influencing hospitalization in BD patients.

The use of ML techniques in psychiatric research, particularly in the context of BD, has the potential to deepen our understanding of the underlying mechanisms and pathophysiology of the condition. By uncovering novel associations and risk factors for patient admissions, these models contribute to the ongoing efforts to unravel the complexities of BD and guide future research directions.

The datasets presented in this article are not readily available due to patient confidentiality and data protection. The raw data supporting the conclusions of this article will be made available by the corresponding author only if this request is approved by the Research Ethics Committee of Fundación Universitaria Sanitas, Bogotá D.C., Colombia, following patient protection regulations in Colombia. Requests to access the datasets should be directed to the corresponding author.

The studies involving humans were approved by Comité de Ética en Investigación Fundación Universitaria Sanitas (CEIFUS). The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and institutional requirements.

MP-A: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing, Project administration. EM-M: Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. JMu: Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. RA-D: Data curation, Investigation, Methodology, Validation, Visualization, Writing – original draft, Writing – review & editing. GL-C: Conceptualization, Data curation, Investigation, Supervision, Visualization, Writing – original draft, Writing – review & editing. AC-J: Conceptualization, Data curation, Formal analysis, Supervision, Visualization, Writing – review & editing. JR-A: Conceptualization, Data curation, Formal analysis, Methodology, Supervision, Validation, Visualization, Writing – review & editing. MA-B: Conceptualization, Investigation, Methodology, Resources, Supervision, Validation, Visualization, Writing – review & editing. JMc: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Visualization, Writing – review & editing.