Development of a machine learning-based depression risk identification tool for older adults with asthma

Abstract

Background

Asthma is a chronic inflammatory disorder that adversely affects the quality of life, particularly in older adults. The coexistence of depression in asthma patients complicates their management and exacerbates health outcomes. This study aims to develop a machine learning-based Depression Risk Identification Tool (DRIT) to predict depression risk in this population.

Methods

We conducted a secondary analysis of data from the China Health and Retirement Longitudinal Study (CHARLS), including 1154 asthma patients. Using LASSO regression, we identified 21 significant predictors of depression. We evaluated eight machine learning algorithms, including the glmBoost model, which was selected based on performance metrics such as accuracy and area under the ROC curve (AUC).

Results

The glmBoost model demonstrated superior predictive performance, achieving an AUC of 0.740 (95% CI: 0.674–0.804) in the testing cohort and 0.664 (95% CI: 0.614–0.714) in the validation cohort. Key risk factors identified included poor cognitive function, heavy exercise, unmarried status, and female gender. The model’s interpretability was enhanced using SHAP values, providing insights into the contributions of each predictor.

Limitations

The study’s reliance on survey data may limit the comprehensiveness of risk factor identification. Additionally, the applicability of findings may vary across different populations, necessitating further validation in diverse cohorts.

Conclusion

The DRIT effectively predicts depression risk among older asthma patients, enabling timely identification and intervention. This tool has the potential to improve patient outcomes and reduce the burden on healthcare systems by facilitating integrated management of asthma and depression.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12888-025-07338-6.

Introduction

Asthma is recognized as a prevalent chronic inflammatory disorder of the airways, significantly diminishing the quality of life for those affected [18]. This impact is particularly acute among older adults populations, who often experience a gradual decline in their physical function, exacerbated by recurrent symptoms such as wheezing and shortness of breath [17]. The persistent nature of these symptoms not only compromises their physiological well-being but also has profound implications for their mental health, leading to an increased risk of psychological disorders, particularly depression [2, 20]. Recent research has highlighted that older adults suffering from asthma face a significantly elevated risk of developing depression compared to the general population [5, 33]. Various factors contribute to this heightened vulnerability, including deterioration in physical capabilities, changes in lifestyle, and inadequate social support systems [8, 28, 37].

Despite the clear connection between asthma and depression, individuals with asthma who exhibit symptoms of depression often do not receive the appropriate treatment [15]. The underlying reasons for this treatment gap are multifaceted. Many patients may lack awareness or harbor misconceptions regarding depression, which can prevent them from seeking help. Additionally, the stigma associated with a depression diagnosis can lead to fears of social isolation, further exacerbating their mental health challenges [12, 23]. Concerns over the potential side effects of antidepressant medications can also deter patients from pursuing necessary treatment [3]. Compounding these issues is the insufficient availability of healthcare professionals qualified to conduct psychological and psychiatric assessments, which contributes to the underdiagnosis of depression among asthma patients [22]. This underdiagnosis is particularly concerning, as depression not only diminishes life satisfaction but can also adversely affect treatment adherence and overall asthma management, creating a detrimental cycle that further burdens both patients and the healthcare system [4, 9].

The implications extend beyond individual health, as the combined effects of asthma and depression impose significant strain on healthcare resources [21, 25]. Consequently, accurately predicting the risk of depression in older asthma patients is essential for facilitating timely interventions, which can enhance both physical and mental health outcomes. In this context, the establishment of an effective depression prediction model is crucial. Such a model could aid healthcare providers in identifying patients at risk for depression, enabling proactive management strategies that address both asthma and its psychological comorbidities.

Traditional methods for screening depression typically rely on clinical evaluations and self-reported questionnaires. While these approaches can yield valuable insights, they often lack the objectivity and timeliness required to capture the dynamic fluctuations in patients’ mental states. The emergence of machine learning technologies in recent years, however, has opened new avenues for healthcare innovation [31, 35]. By leveraging big data analysis, it becomes possible to identify potential risks for depression in a more comprehensive and accurate manner. Machine learning algorithms can analyze vast amounts of multidimensional data, encompassing physical, psychological, and socioeconomic variables, to identify key factors influencing the onset of depression [26, 27].

The primary objective of this study is to develop a machine learning-based depression prediction model specifically designed for older patients with asthma. By systematically gathering and analyzing diverse data sources, we aim to pinpoint critical factors that contribute to the development of depressive symptoms. This research not only seeks to enhance the specificity and effectiveness of clinical interventions but also aims to provide innovative strategies for improving the overall health and well-being of older individuals living with asthma.

By focusing on the intersection of asthma and mental health, this study aspires to bridge the gap in current clinical practice, offering a robust predictive tool that can facilitate early identification of depression risk. Ultimately, our goal is to empower healthcare professionals with the insights necessary to implement timely and targeted interventions, thereby improving both the quality of life for patients and the efficiency of healthcare delivery. The integration of machine learning into the clinical management of asthma and depression represents a promising frontier in personalized medicine, with the potential to transform the landscape of care for vulnerable populations.

Methods

Study population

This research involved a secondary analysis of data sourced from the China Health and Retirement Longitudinal Study (CHARLS). CHARLS (http://charls.pku.edu.cn/) is a national initiative designed to explore health and policy-related information among individuals aged 45 and older, addressing the challenges of an aging population.

Over the period from 2011 to 2018, four surveys were conducted, with participants selected from both rural and urban areas using a multistage stratified sampling method that is proportional to size. Comprehensive details regarding the study’s design and cohort characteristics have been outlined in earlier publications [36].

The data from CHARLS has been extensively utilized in epidemiological research. The study received ethical approval from the Institutional Review Board at Peking University (IRB00001052-11015). Individuals who did not have essential sociodemographic information (such as age and sex), required blood tests and anthropometric data, or information regarding their history of asthma at baseline were also excluded from the study. Finally, a total of 1,154 participants with asthma from CHARLS 2011 and CHARLS 2015 were included in the final analysis.

Research variable

The cognitive and depressive symptoms assessed using the 10-item version of the Center for Epidemiologic Studies Depression Scale (CES-D) [1]. Patients with CES-D score > 10 were considered depressed. A total of 63 potential factors were identified, which included Socio-demographic factors (age, gender, marital status, retire, and education level), lifestyle and health behaviors (smoking, alcohol consumption, social activity participation), laboratory findings (white blood cell count, hemoglobin, hematocrit, mean corpuscular volume, platelet count, triglycerides, creatinine, blood urea nitrogen, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, total cholesterol, glucose, uric acid, cystatin C, C-reactive protein, and glycated hemoglobin), health status (a history of chronic disease (hypertension, dyslipidemia, cancer, heart disease, chronic lung disease, stroke, mental disease, arthritis or rheumatism, liver disease, kidney disease, digestive disease)), economic factors (Per capita household income and consumption). For missing values, we performed multiply-imputed by chained equations to impute covariates through the “mice” package.

Development and evaluation of the depression risk identification tool

In this study, CHARLS 2015 (n = 703) cohort was splitted into a 70% training cohort and a 30% testing cohort. CHARLS 2011 (n = 451) cohort was used for the validation cohort. To develop a depression risk identification tool (DRIT) for older asthma patients, we leveraged machine learning algorithms, comprising The least absolute shrinkage and selection operator (LASSO), elastic-net-regularized generalized linear model (glmNet), bootstrap aggregation classification and regression trees (Bagged CART), Naive Bayes (NB), Partial Least Squares (pls), K-Nearest Neighbors (KNN), random forest (RF), boosted generalized linear model (glmBoost), classification and regression trees (CART). The final signature was generated following such a pipeline: (i) LASSO algorithm was applied to the training cohort to select significant features. (ii) Seven Machine learning algorithms were employed on significant features to individually fit models. In order to mitigate the risk of overfitting resulting from the excessive complexity of the model, a ten fold 10-repeated cross-validation approach was implemented to enhance the generalization capability of the training cohort. (iii) In the testing cohort, a consensus evaluation strategy was applied to assess effectiveness and applicability of all models, encompassing the accuracy, Harrell’s concordance index (C-index), F1-score, precision, and recall. The glmBoost model was recognized as the optimal scheme. Then, the ROC curve was plotted through the glmboost model predictions and the real results. The Youden index was applied to select the threshold. Additionally, decision curve analysis (DCA) was also conducted to assess the effectiveness and relative advantage of each model within the practical context.

Model interpretation

In order to improve better interpretation of glmboost model, we performed the SHapley Additive exPlanation (SHAP) method. We evaluated the importance of each feature by computing the mean absolute value of its SHAP value. Besides, we provided an example of SHAP prediction for the purpose of demonstration.

Results

Characteristics of the study population

A total of 1154 participants from CHARLS 2011 and CHARLS 2015 were recruited in this study. In-depth baseline data for both cohorts were detailed in Table S1. Within the 451 participants from CHARLS 2011 Cohort, 54.77% of them presented with depression. In the CHARLS 2015 cohort, 47.4% of the 703 participants were found to have comorbid depression.

Feature selection

We employed LASSO regression to identify significant indicators associated with the depression. Utilizing the training cohort of 493 patients, we standardized all clinical indicators prior to analysis. Through 10-fold cross-validation, we systematically evaluated a range of λ values, assessing their impact on model performance.

Finally, the optimal lambda value was 0.0334 by using the LASSO algorithm and 21 indicators were selected, including gender, marital status, activity of daily living (ADL), coronary heart disease, stroke, arthritis, dyslipidemia, liver disease, kidney disease, digestive disease, poor memory, heavy exercise, hospitalizations in the past year, retirement, drop around, play mahjong, exercise, cognition, hematocrit, and blood urea nitrogen (Fig. 1A-B).

Fig. 1.

Variable selection through the LASSO regression algorithm. A In the LASSO regression model, we select the best parameter (λ). In this plot, the horizontal axis represents log λ, while the vertical axis shows the regression coefficients. B Another plot illustrates the relationship between λ and the number of variables. Here, the bottom horizontal axis is log λ, the top horizontal axis indicates the number of variables, and the vertical axis represents binomial deviance

Machine learningbased integrative program generates the depression risk identification tool and evaluation

21 identified significant indicators were incorporated into our machine learning-based integrative program to establish the depression risk identification tool (DRIT). We employed seven classical learners using tenfold cross-validation with 10 repetitions to fit models and evaluated their performance using five metrics, including accuracy, C-index, F1-score, precision, recall in the testing cohort. The boosted generalized linear model demonstrated superior classification capabilities based on these performance metrics (Fig. 2A, Table S2).

Fig. 2.

Machine learning-based integrative program generates DRIT. A Comprehensive performances of seven types of algorithms. Circles showed the distribution of recall, precision, F1-score, C-index, and accuracy of each algorithm

Robust performance of DRIT

To verify the accuracy and universality of the diagnostic model, receiver-operator characteristic (ROC) analysis was used to preliminarily measure the discrimination of the model, the area under the ROC curve (AUC) [95% confidence interval] of the diagnostic model in the test and validation cohort was 0.740 (95%CI: 0.674–0.804), and 0.664 (95%CI :0.614–0.714), respectively (Fig. 3A-B).

Fig. 3.

Comprehensive evaluation of DRIT. ROC curves for the testing (A) and validation (B) cohort; confusion matrices for the testing (C) and validation (D) cohort; the decision curve analysis of DRIT for depression in older asthma individuals within the testing (E) and validation cohort (F)

In the test and validation cohort, The confusion matrix provides detailed discriminant analysis results. The higher precision, specificity, kappa, F1 score, accuracy and AUC consistently indicate that DRIT generated through the glmBoost model classifier stably distinguish individuals with and without depression among older asthma patients, which has high accuracy and excellent performance (Fig. 3C-D). Then, to assess DRIT’s practical utility, we used decision curve analysis to plot curves on the test and validation cohort. For the optimal decision threshold > 0%, the glmboost model showed significant benefit for clinical intervention (Fig. 3E-F).

Model interpretation for DRIT

To better understand the relationship between the model and the data, we provide a more intuitive interpretation of the best-performing glmBoost model using SHAP to illustrate how these variables affect the depression rate in the model. Figure 4A showed importance ranking of 15 evaluated factors by mean absolute SHAP values, demonstrated their respective contribution to model. Additionally, we provided a typical prediction for depression examples to demonstrate the model’s interpretability. As demonstrated, poor cognition, heavy exercise, unmarried, and female asthma patients tended to be depressed (Fig. 4B).

Fig. 4.

Model interpretation for the DRIT. A Ranking of variable importance based on the average value. B SHAP predictions related to depression show yellow arrows for higher risk and purple arrows for lower risk. The length of the arrows helps us see how much influence each feature has; the longer the arrow, the more important that feature is for the outcome

Discussion

Asthma is a chronic inflammatory disease of the airways, characterized by airway hyperresponsiveness that leads to recurrent episodes of wheezing, dyspnea, chest tightness, and/or coughing [24]. When not adequately controlled, these symptoms can severely hinder daily activities and diminish overall quality of life [19]. A notable complication associated with asthma is the increased prevalence of depression, which complicates the management of both conditions. Previous research indicates that individuals with asthma who also suffer from depression are 3.17 times more likely to experience exacerbations of their asthma compared to those without depression. Furthermore, patients with asthma have been found to be 1.52 times more likely to present with additional comorbidities compared to those without asthma, highlighting the interconnected nature of these health issues [10]. The presence of depression not only exacerbates asthma symptoms but also negatively affects treatment adherence, which can create a vicious cycle of worsening health outcomes [7]. This situation places a considerable burden on patients, both in terms of their health and the economic implications, while also presenting challenges for healthcare providers.

The relationship between asthma and depression is intricate, with significant alterations in cytokine levels observed in both conditions. Some researchers propose that allergy-related inflammatory responses may play a crucial role in the shared pathophysiology of asthma and depression [30]. Specific biomarkers have also been identified as potential links between the two disorders. For instance, the NOD-like receptor family pyrin domain-containing 3 (NLRP3) is believed to activate the hypothalamic-pituitary-adrenal (HPA) axis through the elevation of serum levels of pro-inflammatory cytokines, including IL-1β, IL-6, and TNF-α. This activation ultimately leads to increased glucocorticoid release in patients with asthma [14]. Prolonged exposure to elevated glucocorticoid levels can result in neuronal atrophy, inhibit neurogenesis in critical brain regions such as the hippocampus and prefrontal cortex, and decrease synaptic plasticity. Additionally, overexpression of IL-1β can further diminish brain-derived neurotrophic factor (BDNF) production and neurogenesis in the hippocampus, which may collectively predispose individuals to depressive symptoms [29]. Moreover, the association between NLRP3 and pyroptosis suggests a potential role of pyroptosis in the comorbidity of asthma and depression, indicating a need for further investigation into this relationship [11, 32]. Despite the recognition of these complexities, current clinical strategies to address the dual diagnosis of asthma and depression lack effective, universally accepted intervention [6]. The challenging circumstances underscore the urgent need to equip clinicians with reliable tools for early identification of high-risk asthma patients and timely intervention, ultimately aiming to reduce the prevalence of depression.

In response to this critical need, we have developed a machine learning-based prediction system designed to forecast the risk of depression in older patients with asthma. The selection of features or variables is a crucial aspect of predictive model development. Starting with an initial pool of 63 variables, we employed the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm, which enabled us to isolate 21 significant variables that made meaningful contributions to the predictive model. Following this, we conducted a comprehensive evaluation of seven distinct machine learning models, assessing them based on various metrics including accuracy, C-index, F1-score, precision, and recall. Ultimately, we identified the optimal Depression Risk Identification Tool (DRIT) through the glmBoost algorithm, which demonstrated superior predictive performance, achieving an area under the curve (AUC) of 0.740 (95% CI: 0.674–0.804). The robustness of the DRIT was confirmed through external validation, yielding an AUC of 0.664 (95% CI: 0.614–0.714). The glmBoost model has garnered significant attention in the realm of clinical prediction research due to its rapid computation speed, strong generalizability, and high predictive accuracy. Furthermore, our analysis revealed the DRIT’s capacity for accurate discrimination, robustness, and clinical net benefit in identifying patients with comorbid depression among older asthma patients, as demonstrated through Receiver Operating Characteristic (ROC) analysis, confusion matrix evaluations, and Decision Curve Analysis (DCA).

Traditional machine learning algorithms have often led to confusion among users due to their inherent lack of transparency and interpretability. To address this issue, our study integrates the use of SHAP (SHapley Additive exPlanations) values to quantify risk factors, thereby providing clinicians with an intuitive understanding of the characteristics that influence the susceptibility of asthma patients to depression. The top five features ranked by their importance score (in descending order) included age, cognitive ability, levels of physical exercise, marital status, and gender. Specifically, older asthma patients who exhibited poor cognitive function, engaged in heavy exercise, were unmarried, and female demonstrated a heightened risk of developing depression. Conversely, older asthma patients (aged over 79) were found to possess a relatively lower risk of experiencing depressive symptoms. Numerous studies have established a correlation between cognitive deficits and depression in older adults [13, 16]. Additionally, large-scale studies involving populations from seven different countries have confirmed that unmarried individuals are at a greater risk of depression [34]. Therefore, we posit that these characteristics play a significant role in the onset and progression of depression. We recommend that clinicians consider these factors and use the Depression Risk Identification Tool (DRIT) to promptly identify older individuals with asthma who are at high risk for depression. This proactive approach will enable them to provide timely psychological therapy and medication.

In conclusion, our study has yielded a reliable Depression Risk Identification Tool (DRIT) that effectively predicts the risk of depression in older patients suffering from asthma. By equipping clinicians with accurate risk assessment tools, we aim to facilitate the prompt identification and intervention for high-risk individuals, thereby minimizing the incidence of depression. This advancement holds considerable promise for enhancing patient outcomes and alleviating the burdens faced by both patients and healthcare providers.

However, it is important to acknowledge the limitations of this study. First, the variables chosen were constrained by the design of the survey questionnaire, which means we cannot guarantee that all relevant risk factors have been adequately captured. For instance, the absence of the asthma control status variable affects our understanding of the relationship between asthma and depression. The decision regarding whether to categorize the selected variables as categorical or continuous, along with the criteria employed for making such divisions, may also influence the findings and their interpretations. Exploring a continuum of CES-D scores was a better way to capture the range of depressive symptoms in future. Second, the data utilized in this study were obtained from a nationally representative survey, which could potentially limit the applicability and benefit thresholds in specific regions. Besides, self-reported assessments can be influenced by various factors, including the respondents’ awareness and understanding of their mental health status, as well as potential stigma associated with depression. Furthermore, while the DRIT shows promise, it requires validation in more externally independent cohorts to further assess its reliability and generalizability across diverse populations. Despite these limitations, we firmly believe that our tool offers valuable support for the prevention and management of depression among older adults with asthma, paving the way for improved clinical outcomes and enhanced quality of life for this vulnerable population.

Conclusion

Asthma significantly affects the quality of life, particularly in older individuals, and is closely linked to an increased risk of depression. This study introduces the Depression Risk Identification Tool (DRIT), a machine learning-based model that accurately predicts depression risk using diverse factors. The glmBoost model demonstrated superior predictive performance, identifying critical risk factors such as cognitive ability and marital status. Despite limitations in data and the need for further validation, the DRIT offers a valuable resource for healthcare providers to facilitate early interventions, ultimately enhancing patient outcomes in this vulnerable population.

Authors’ contributions

L.A. and R.W. contributed study design and paper revising. X.W and L.J. contributed project oversight. M.L. and H.W. contributed data analysis, visualization, and paper writing. All authors approved this manuscript.

Funding

The authors did not receive any financial support for this study.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was conducted following the principles of the Helsinki Declaration and was approved by the Biomedical Ethics Committee of Peking University. All participants signed informed consent forms before their participation, which were approved by the Ethics Review Committee of Peking University (IRB00001052-11015).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.