Machine learning-driven simplification of the hypomania checklist-32 for adolescent: a feature selection approach Authors: Guanghui Shen, Haoran Chen, Xinwu Ye, Xiaodong Xue, Shusi Tang Published: 2024 ABSTRACT AbstractBackgroundThe Hypomania Checklist-32 is widely used to screen for bipolar disorder, but its length can be challenging for adolescents with manic symptoms. This study aimed to develop a shortened version of the HCL-32 tailored for adolescents using machine learning techniques.MethodsData from 2,850 adolescents (mean age 15.50 years, 68.81% female) who completed the HCL-32 were analyzed. Random forest (RF) and gradient boosting machine (GBM) algorithms were employed for feature selection. The area under the curve (AUC) was used to evaluate model performance. Receiver operating characteristic (ROC) analysis was conducted to determine optimal cutoff points for the shortened scale.ResultsAn 8-item version of the HCL-32 was derived, maintaining high predictive accuracy (AUC = 0.97). The selected items captured core symptoms of adolescent mania, including increased energy, risk-taking, and irritability. Two cutoff points were identified: a score of 3 offered high specificity (0.98) and positive predictive value (0.98), while a score of 4 provided balanced sensitivity (0.87) and specificity (0.94) with the highest overall accuracy (0.91).ConclusionsThe machine learning-driven 8-item version of the HCL-32 demonstrates strong diagnostic utility for adolescent bipolar disorder, offering a more efficient screening tool without sacrificing clinical sensitivity. This shortened scale may improve assessment feasibility and accuracy in clinical settings, addressing the unique challenges of diagnosing bipolar disorder in adolescents.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40345-024-00365-4.Keywords: Feature selection, Random forest, Gradient boosting machine, Scale reduction, Manic symptoms Introduction Bipolar disorder, characterized by alternating episodes of mania and depression, represents a significant challenge in adolescent psychiatry. The prevalence of bipolar disorders in adolescents is estimated to be between 1.8 and 2.5% (Gautam et al. 2019; Parry et al. 2018), with the onset frequently occurring during this critical developmental period (15–25 years) (Meter et al. 2019). Adolescent mania, a key feature of bipolar disorder in this age group, presents unique diagnostic and therapeutic challenges due to its complex symptomatology and potential overlap with other psychiatric conditions common in adolescence (Duffy et al. 2020). The clinical presentation of mania in adolescents is often characterized by marked irritability, emotional lability, and atypical symptoms that may not fully align with adult diagnostic criteria (Leibenluft et al. 2003; Ryles et al. 2017). For instance, a longitudinal study by Findling et al. (2018) revealed that adolescents with bipolar disorder experienced more severe irritability and had poorer global functioning compared to adults with the same diagnosis, underscoring the unique presentation in this age group (Findling et al. 2018). Notably, inattention and irritability emerge as core features of adolescent mania, complicating the diagnostic process and necessitating specialized assessment approaches (Comparelli et al. 2022). Research by Sandstrom et al. (2021) showed that up to 70% of youth with bipolar disorder experience significant attention deficits during manic episodes (Sandstrom et al. 2021). These symptoms not only impact the individual’s functioning but also pose significant challenges in clinical assessment and research methodologies. Given the complexity of diagnosing adolescent mania, clinicians often rely on standardized assessment tools to aid in identifying manic symptoms. The Hypomania Checklist (HCL-32) is one such tool, widely recognized for its ability to capture a broad spectrum of hypomanic symptoms (Wang et al. 2022). Developed to screen for bipolar disorder, the HCL-32 consists of 32 items that inquire about various aspects of mood, behavior, and cognitive functioning (Angst et al. 2005). The HCL-32 is particularly valued for its sensitivity in detecting hypomanic symptoms, which are often underreported or misinterpreted in clinical settings (Lee et al. 2016). However, despite its potential utility, the application of the HCL-32 in adolescent populations, particularly those experiencing manic symptoms, presents significant challenges. In clinical practice, the length of the HCL-32 often proves to be a substantial barrier to its successful completion by adolescent patients experiencing manic symptoms, and researchers have begun to focus on simplification (Wang et al. 2019). This challenge is exacerbated by the core features of adolescent mania previously discussed—inattention and irritability (Evans et al. 2017), making it difficult for them to maintain focus throughout a 32-item questionnaire. Furthermore, the heightened irritability characteristic of adolescent mania can lead to frustration and disengagement during lengthy assessment procedures (Brotman et al. 2017). This result in incomplete questionnaires or refusal to participate. Feature selection techniques in machine learning offer a promising approach to address the challenges associated with the HCL-32 in adolescent populations (Li et al. 2017). These data-driven methods can identify the most informative items from the original items scale (Ali et al. 2018), potentially creating a shorter, more focused version that maintains clinical sensitivity while improving convenience and completion rates among adolescents experiencing manic symptoms. Unlike traditional item reduction methods, machine learning algorithms can capture complex, non-linear relationships between items, which is particularly relevant in psychological assessments where symptoms interact in intricate ways (Liang et al. 2022). Feature selection, a core aspect of machine learning, allows us to systematically identify the most informative items from the scale, reducing redundancy and ensuring that only the most clinically relevant factors are retained (Li et al. 2017). This process is especially beneficial in populations such as adolescents with bipolar disorder, where cognitive and emotional challenges—such as inattention (Kent and Craddock 2003) and irritability (Stringaris 2011)—can reduce the feasibility of completing long questionnaires. The purpose of this research is to develop a more efficient version of HCL-32 tailored to adolescents with bipolar disorder, utilizing machine learning techniques to compress the scale while maintaining its clinical sensitivity and diagnostic utility. Given the unique diagnostic challenges posed by adolescent mania, including inattention and irritability, a shorter and more focused tool would improve patient compliance and the accuracy of assessments. By ensuring that the most relevant items are retained through machine learning-based feature selection, the resulting scale will offer a streamlined, yet highly effective, approach to diagnosing and managing bipolar disorder in adolescents. Method Participants Participants for this study were recruited from the outpatient adolescent population of the Child Psychiatry Department at Wenzhou Seventh People’s Hospital between June 2023 and June 2024. Eligibility criteria included completion of the HCL-32 assessment as part of the outpatient diagnosis and treatment process, and an age range of 12 to 18 years at the time of assessment. The clinical diagnoses in this study (manic or bipolar disorder), were made by two licensed psychiatrists (at least one holding a senior professional title) according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-V) criteria. Exclusion criteria were applied to ensure a focused study population, eliminating individuals with a history of organic disease, mental retardation, or brain trauma. Additionally, to maintain data quality, participants who provided regular pattern answers (e.g., alternating yes/no responses) on the HCL-32 were also excluded from the analysis. Participants were grouped into two categories based on their clinical diagnosis and HCL-32 results. The “With manic symptoms” group included those diagnosed with bipolar disorder or mania and HCL-32 score above the cut-off point. The “Without manic symptoms” group consisted of individuals who did not meet these criteria. All eligible participants, along with their parents or legal guardians for those under 18, provided informed consent prior to study involvement. The study protocol received approval from the Ethics Committee of Wenzhou Seventh People’s Hospital (No. 202402). The final sample consisted of 2850 participants, with a mean age of 15.50 years (SD = 1.73). The gender distribution was 68.81% female and 31.19% male. Measurements Demographic information Demographic data were extracted from the outpatient system, including gender and age. Gender was transformed into a dummy variable, with male participants coded as 0 and female participants coded as 1. Age was also converted into dummy variables to facilitate analysis: participants aged 16 years and below were coded as 0, while those aged 17–18 were coded as 1. Manic symptoms Manic symptoms were assessed using the Hypomania Checklist-32 (HCL-32), a self-report questionnaire designed to screen for hypomanic symptoms (Angst et al. 2005). The HCL-32 consists of 32 yes/no items that inquire about various aspects of mood, behavior, and cognitive functioning typically associated with hypomanic episodes. The HCL-32 has demonstrated good psychometric properties in both adult and adolescent populations. It has a well-established cut-off point for identifying individuals likely to have experienced hypomanic episodes. In the original validation study, a cut-off score of 14 or more positive responses yielded the best combination of sensitivity (80%) and specificity (51%) for distinguishing between bipolar and unipolar depression in an adult population. In adolescent populations specifically, the HCL-32 has demonstrated promising psychometric properties, with Cronbach’s α = 0.87 in current study. Statistical analyses All statistical analyses were conducted using R software (version 4.4.1), with significance levels set at p < 0.05 unless otherwise specified. The overall analytical approach aimed to assess the diagnostic utility of a compressed version of the HCL-32 by employing machine learning techniques for feature selection and evaluating model performance based on multiple classification metrics. Descriptive statistics Descriptive statistics were first computed to summarize the demographic characteristics of the participants. Continuous variables, such as age, were reported as means with standard deviations (SD), while categorical variables, such as gender, were expressed as frequencies and percentages. Chi-square tests were employed to examine the associations between categorical demographic variables (e.g., gender, age) and the presence of manic symptoms as assessed by the HCL-32. Feature selection and machine learning algorithms Random forest (RF) (Feng et al. 2022) and gradient boosting machine (GBM) (Adler and Painsky 2022) algorithms were employed to perform feature selection, identifying the most predictive items from the original 32-item HCL scale. Both algorithms were applied using a hold-out validation approach, with 70% of the dataset used for training and 30% reserved for testing, to minimize overfitting and ensure robust model performance. The importance of each item was ranked based on the mean decrease in accuracy (MDA) for RF and the Gain metric for GBM. The area under the receiver operating characteristic curve (AUC) was used as the primary metric to evaluate the classification performance of the selected features. A sequential feature selection method was applied, with AUC values computed for models including an increasing number of features (from 1 to 32) to identify the optimal number of features for maximal predictive accuracy. The final model was selected based on the point where AUC stabilized, indicating that additional features did not significantly enhance performance. Model evaluation Receiver operating characteristic (ROC) curves were generated to assess the discriminative power of the compressed HCL-32 scale in predicting manic symptoms. The AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated for the compressed HCL-32 scale. Confusion matrices were used to visualize classification performance at various cutoff points. All machine learning analyses were performed using the “ranger” and “gbm” packages in R. ROC analyses were conducted using the “pROC” package. Result Descriptive analysis The demographic characteristics of the participants, along with their distribution based on the presence or absence of manic symptoms, are summarized in Table 1. Based on the clinical diagnosis and the HCL-32 assessment, 1580 (55.44%) participants were classified as having manic symptoms, while 1270 (44.56%) did not meet the threshold for manic symptoms. The mean HCL-32 score for the entire sample was 13.31 (SD = 6.67). The result of chi-square tests showed that no significant association between gender and manic symptoms (χ2 = 0.62, p = 0.43), and similarly, no significant relationship was found between age group and manic symptoms (χ2 = 1.25, p = 0.26). Table 1Demographic information (N = 2850)AllWith manic symptoms (N = 1580)Without manic symptoms (N = 1270)χ2pGender0.620.43 Male889(31.19%)503(56.58%)386(43.42%) Female1961(68.81%)1077(54.92%)884(45.08%)Age (years)1.250.26 Below and 161933(67.83%)1086(56.18%)847(43.82%) 17–18917(32.17%)494(53.87%)423(46.13%) Demographic information (N = 2850) Feature extraction Random forest and gradient boosting machine (GBM) algorithms were employed for feature extraction, with the area under the curve (AUC) serving as the performance metric. Item2, item12 consistently emerged as the most important features across both algorithms. For a full comparison of feature rankings between random forest and GBM, see Fig. 1. As shown in the Fig. 2 and Table 2, the AUC increased as the number of selected features increased. However, after reaching 8 features (AUC = 0.97), the AUC began to stabilize, the additional features did not significantly enhance the model’s predictive performance [Full iteration were presented in Supplementary Materials 1(RF) and 2(GBM)]. The combination of 8 features identified as the most important for predicting manic symptoms were as follows: Item 2 (“I feel more energetic and more active”), Item 5 (“I am more sociable (make more phone calls, go out more)”), Item 9 (“I take more risks in my daily life (in my work and/or other activities)”), Item 12 (“I have more ideas, I am more creative”), Item 17 (“I am more flirtatious and/or am more sexually active”), Item 18 (“I talk more”), Item 20 (“I make more jokes or puns when I am talking”), and Item 27 (“I get into more quarrels”). For more detailed data, refer to Supplementary Material 1. Feature importance ranking for predicting manic symptoms by random forest (A) and gradient boosting machine (B) AUC under different feature combinations. Random forest (Left) and gradient boosting machine (Right) Table 2Comparison of average and highest AUC for random forest (RF) and gradient boosting machines (GBM) across different numbers of factorsNumber of factorsRF. Average AUCRF. Highest AUCGBM. Average AUCGBM. Highest AUC10.670.810.670.8220.860.890.860.8930.910.920.910.9340.930.940.940.9550.940.950.960.9660.950.960.960.9770.960.960.970.9780.970.970.970.9890.970.970.980.98100.970.970.980.98…320.990.990.990.99 Comparison of average and highest AUC for random forest (RF) and gradient boosting machines (GBM) across different numbers of factors Model evaluation Following the selection of 8 key factors, the total score for these factors was calculated and subjected to a Receiver Operating Characteristic (ROC) analysis. The results indicated that the optimal cutoff value was 3.50, with an AUC of 0.96, indicating excellent discriminative ability. However, for practical application in clinical settings, it is necessary to use integer cutoff values. Therefore, further ROC analysis and confusion matrix evaluations were conducted at cut points 3 and 4. As shown in Figs. 3, 4 and Table 3, both cutoff points demonstrate strong performance (AUC = 0.83, AUC = 0.90), but with different trade-offs. Cutoff point 3 offers extremely high specificity (0.98) and positive predictive value (0.98), making it excellent for ruling in manic symptoms, but with lower sensitivity (0.68), potentially missing some cases. Its overall accuracy is 0.82. In contrast, cutoff point 4 provides a more balanced performance with higher sensitivity (0.87) while maintaining high specificity (0.94). Both its positive and negative predictive values are high (0.94 and 0.86, respectively), achieving the highest overall accuracy of 0.91. The confusion matrices (Figs. 3 and 4) visually represent these performance characteristics. RoC at cut point 3 (red) and cut point 4 (black) for the selected 8 factors Performance metrics comparison for different cutoff points of the 8-item HCL scale Table 3Comparison of performance metrics for different cut off points of the 8-items HCLOptimal cutoff pointSensitivitySpecificityPositive predictive valueNegative predictive valueAccuracyCut point 30.680.980.980.710.82Cut point 40.870.940.940.860.91 Comparison of performance metrics for different cut off points of the 8-items HCL Discussion This study aimed to compress the HCL-32 into a shorter, more clinically practical version tailored for adolescents with bipolar disorder, using feature selection techniques from machine learning. By RF and GBM algorithms, we identified an optimal set of 8 items that retained the diagnostic sensitivity and specificity of the original scale while improving convenience and feasibility in clinical settings. The reduced scale demonstrates strong predictive accuracy, with an AUC of 0.91, confirming its potential utility in adolescent psychiatry. Our results reveal that the shortened 8-item version of the HCL-32 performs comparably to the full 32-item scale, achieving high AUC values across both RF and GBM models. The high AUC scores indicate that this reduced scale maintains its diagnostic accuracy while offering a more feasible tool for clinical use, especially in time-constrained or high-stress environments where adolescents with manic symptoms may struggle to complete longer assessments. These findings underscore the potential for feature selection algorithms to reduce the length of psychiatric assessment tools without sacrificing clinical sensitivity, which is consistent with the growing body of research applying machine learning to mental health diagnostics (Jacobucci and Grimm 2020; Shatte et al. 2019). The selected items, such as “I feel more energetic” and “I take more risks,” capture the core behavioral and cognitive features of mania in adolescents. These symptoms—associated with increased energy, risk-taking behaviors, and impulsivity—are characteristic of adolescent mania, which often presents with unique challenges compared to adult presentations of the disorder. Machado-Vieira et al. (2017) found that increased energy and goal-directed activity were among the most common symptoms in adolescents with bipolar disorder (Machado-Vieira et al. 2017). Similarly, Ng et al. (2016) highlighted the prevalence of risk-taking behaviors and impulsivity in manic episodes among youth (Ng et al. 2016). Our study adds to this literature by identifying a minimal set of items that can effectively capture these core symptoms, thus providing clinicians with a streamlined yet effective tool for diagnosing mania in adolescents. Interestingly, our machine learning approach also identified Item 27 (“I get into more quarrels”) as a crucial feature. This aligns with the growing recognition of irritability as a central feature of pediatric bipolar disorder, a concept explored in depth by Leibenluft et al. (2011). The prominence of this item in our compressed scale underscores the importance of considering age-specific manifestations of manic symptoms in diagnostic tools. This study supports the use of data-driven feature selection methods in psychiatric assessment, offering an innovative approach to scale development. Traditional item reduction methods often rely on factor analysis or expert judgment, which may overlook complex non-linear relationships between symptoms (Bystritsky et al. 2012). In contrast, machine learning algorithms can identify such patterns, allowing for more precise and clinically relevant item selection. RF are particularly adept at handling high-dimensional data and capturing complex interactions between features (Reif et al. 2006), while GBM excels at identifying the most influential predictors through its iterative, gradient-based approach (Ji et al. 2022). The consistency in feature importance rankings between these two algorithms lends credibility to our findings and suggests that the selected items are truly the most informative for identifying manic symptoms in adolescents. This multi-algorithm approach addresses concerns raised by Dwyer et al. (2018) about the potential instability of feature selection in psychiatric research (Dwyer et al. 2018). Our findings align with those of Obermeyer and Emanuel, who argue that machine learning techniques hold significant promise for enhancing the diagnostic process across various medical fields, including psychiatry (Obermeyer and Emanuel 2016). The reduced HCL-32 scale not only minimizes the burden on adolescents, who often struggle with attention deficits and irritability during manic episodes, but also enhances clinical workflow by providing a more efficient diagnostic tool. Adolescents with bipolar disorder are known to have difficulty completing lengthy assessments due to inattention (Pavuluri et al. 2006), and a shorter, targeted version of the HCL-32 is likely to improve completion rates and diagnostic accuracy in this population. Moreover, the performance of the reduced scale at different cutoff points highlights its versatility for clinical use. Cutoff point 3 offers high specificity and positive predictive value, making it particularly useful for ruling in manic symptoms with minimal false positives. In contrast, cutoff point 4 provides a more balanced sensitivity, making it a strong candidate for general screening purposes in outpatient settings. Limitation Despite the promising findings, this study has several limitations. First, as a retrospective study utilizing hospital electronic medical records, our methodology has inherent limitations in terms of data collection and standardization. While the clinical diagnoses were made by qualified psychiatrists according to DSM-V criteria during routine care, the retrospective nature of the study means that we could not implement prospective diagnostic protocols or structured clinical interviews. Future prospective studies with standardized diagnostic protocols and real-time data collection would be valuable to address these limitations. Secondly, the dataset was derived from a single clinical population, which may limit the generalizability of the results. Future studies should include diverse populations from multiple sites to validate the robustness of the 8-item HCL-32 across different clinical contexts. Additionally, while the machine learning algorithms performed well in this study, it is essential to replicate these findings using other feature selection methods to ensure the reliability of the item selection process. Finally, this study primarily focused on optimizing the diagnostic utility of the HCL-32, without addressing its prognostic or longitudinal use in monitoring manic symptoms over time. Future research could explore the utility of the reduced scale in tracking symptom changes and treatment outcomes, which would provide a more comprehensive tool for clinicians. Supplementary Information Supplementary Material 1. Supplementary Material 2. Acknowledgements Not applicable. Declarations All procedures were performed in accordance with relevant guidelines from declaration of Helsinki statement. Author contributions Guanghui Shen, Xinwu Ye: formal analysis, writing—original draft, writing—review and editing. Haoran Chen and Xiaodong Xue: writing—review and editing, investigation, conceptualization. Shusi Tang: investigation, methodology, writing—original draft. Funding This work was supported by the Wenzhou Science and Technology Project (Grant No. Y20240046, Y2023865) and Zhejiang Provincial Basic Public Welfare Research (LGF21H090003). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Availability of data and materials The datasets presented in this article are not readily available due to the inclusion of student-specific information. Requests to access the datasets should be directed to shenguanghuipsy@163.com or yexinwu@163.com. Declarations Ethics approval and consent to participate The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All procedures involving human subjects/patients were approved by IRB in Wenzhou Seventh People’s Hospital (No. 202402). Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Guanghui Shen and Haoran Chen have contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Data Availability Statement The datasets presented in this article are not readily available due to the inclusion of student-specific information. Requests to access the datasets should be directed to shenguanghuipsy@163.com or yexinwu@163.com.