Study design
We conducted a cross-sectional study to identify risk factors of psychological distress and develop a prediction model for high-risk psychological distress in lung cancer patients. Optimal predictors were selected to build the prediction model, and the Strengthening the Reporting of Observational Studies in Epidemiology guidelines were followed for reporting results(von Elm et al., 2014).
Study population
We included 342 lung cancer patients at two hospitals in China between June and November 2023 according to the following criteria: (1) cytological or histological diagnosis of lung cancer, (2) age ≥ 18 years, (3) awareness of their diagnosis and willingness to participate, and (4) literacy in Chinese. Exclusion criteria were: (1) concurrent diagnosis of other cancers or psychiatric disorders, (2) history of suicide attempts, (3) participation in other psycho-social interventions within 3 months prior to recruitment, and (4) physical or cognitive impairments (< 26 on the Mini-Mental State Examination) that hindered questionnaire response. The study protocol was approved by the Internal Review Board of Chongqing Traditional Chinese Medicine Hospital and adhered to the Declaration of Helsinki.
Study variables
This study sought to identify the most optimal predictors of psychological distress in lung cancer patients across demographic, clinical, and psycho-social aspects. Demographic variables included age, sex, educational level, employment status, marital status, payment method, residence, number of children, household income, family history of lung cancer, and smoking history. Clinical variables encompassed diagnosis duration, metastasis, comorbidity, pain degree, and TNM stage. Additionally, psycho-social variables were perceived stigma, illness perception, social support, mindfulness, coping styles, perceived stress, and self-esteem.
Study instruments
General information questionnaire
The general information questionnaire was developed by ourselves to collect demographic and clinical data.
Distress thermometer
Psychological distress, the primary outcome variable, was assessed using the Distress Thermometer (DT), a validated single-item measure with an 11-point scale ranging from 0 (no distress) to 10 (extreme distress) visually represented as a thermometer(Riba et al., 2019). The psychometric properties of DT are well-established across diverse populations(Bui et al., 2005; Hong et al., 2015), with several studies consistently suggesting a score of ≥ 4 as indicative of significant distress(Donovan et al., 2014; Tang et al., 2011). Notably, a study demonstrated the applicability of DT ≥ 4 for Chinese cancer population, reporting an area under the curve (AUC) of 0.885(Hong et al., 2015).
Cataldo Lung Cancer Stigma Scale
The 31-item Cataldo Lung Cancer Stigma Scale (CLCSS) was used to assess perceived stigma, which was categorized into four sub-scales: stigma and shame, social isolation, discrimination, and smoking. Items were rated on a 0–4 point scale, with higher scores indicating higher perceived stigma. The CLCSS has been translated into Chinese and demonstrated good psychometric properties, with a Cronbach’s α of 0.932 for the overall scale(Cataldo et al., 2011; YU et al., 2017).
Brief Illness Perception Questionnaire
The Brief Illness Perception Questionnaire (B-IPQ) was developed by Weinman et al.(Weinman and KJ) to assess illness perception on a continuous 0–10 point scale. Higher scores indicate a more negative illness perception. The B-IPQ has demonstrated good test-retest reliability, predictive validity, and discriminant validity(Broadbent et al., 2006). The Chinese version of B-IPQ has also been developed and has been widely used in China(Zhang et al., 2017).
Multidimensional Scale of Perceived Social Support
We employed the 12-item Multidimensional Scale of Perceived Social Support (MSPSS) to assess social support from three sources: family, friends, and significant others. Participants responded on a 7-point scale (1 = strongly disagree, 7 = strongly agree)(Zimet et al., 1990). Prior research has established the psychometric properties of the MSPSS, with sub-scale coefficient alpha values ranging from 0.81 to 0.98(Zimet et al., 1990). The Chinese version of the MSPSS has also demonstrated good psychometric properties in cancer patients(Yang et al., 2009).
Five Facet Mindfulness Questionnaire
The Five Facet Mindfulness Questionnaire (FFMQ), a 39-item instrument developed by Baer et al.(Baer et al., 2006), assesses mindfulness across five facets: observing, describing, acting with awareness, non-judging, and non-reacting. The FFMQ has been translated into Chinese with good psychometric properties(Deng et al., 2011), and a recent study using Bayesian SEM confirmed its psychometric properties in cancer patients(Fong et al., 2021).
Medical Coping Modes Questionnaire
The Medical Coping Modes Questionnaire (MCMQ), a 20-item instrument, was used to assess participants’ coping styles across three sub-scales: confrontation, avoidance, and giving up(Feifel and Strack, 1989). Responses were recorded on a 4-point scale ranging from 1 (never) to 4 (always). The Chinese version of the MCMQ has been validated by Shen et al.(Shen and Jiang, 2000), with Cronbach’s α coefficients exceeding 0.60 for all three sub-scales.
Perceived Stress Scale
Perceived stress was measured using the 10-item Perceived Stress Scale (PSS)(Cohen et al., 1983). Respondents rated each item on a 5-point scale from 0 (never) to 4 (very often), with higher scores indicating higher perceived stress. The Chinese version of the PSS has been validated and demonstrated good internal consistency, with Cronbach’s α of ranging from 0.619 to 0.76(Yuan and Lin, 2009).
Rosenberg Self-Esteem Scale
The Rosenberg Self-Esteem Scale (RSES) is a widely used 10-item instrument for measuring global self-esteem(Rosenberg et al., 1989), with higher scores indicating higher self-esteem. The Chinese version of the RSES was validated by Wu et al.(Wu et al., 2017) and has been extensively used in China(Qian and Yu, 2022; Xiao and Zheng, 2022).
Data collection
A convenience sample of lung cancer patients admitted for treatment at two tertiary hospitals in Chongqing of China between June and November 2023 was screened for eligibility. Eligible patients were approached within 48 hours of admission and provided a detailed explanation of the study’s aims, procedures, and their rights as participants. All participants were assured of anonymity and that collected data would be solely for academic purposes. Written or oral informed consent was obtained before proceeding with a previously piloted questionnaire survey. Patients then self-completed the questionnaires face-to-face in the wards.
Feature selection
Potential variables were selected through a feature selection procedure utilizing the least absolute shrinkage and selection operator (LASSO) regression which is one of the most popular techniques for feature selection(Vasquez et al., 2016). LASSO produces a function of penalty that compresses some of the regression coefficients. Specifically speaking, LASSO analysis pushes the sum of the coefficients’ absolute values to be less than the given value while setting some regression coefficients to zero, thereby producing a more refined model. LASSO retains the advantage of subset shrinkage as a biased estimator for data with complex covariance. This algorithm eliminates factors with zero coefficients automatically using LassoCV, a 10-fold cross-validation approach.
ML model establishment and development
Eight ML algorithms were employed to develop and compare prediction models for psychological distress risk in lung cancer patients. The employed algorithms for model construction included LR, support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), Gaussian naive Bayes (GNB), and adaptive boosting (AdaBoost). Python 3.7 was used to implement all models, with the "xgboost 1.2.1" package for XGBoost, "lightgbm 3.2.1" package for LightGBM, and the "sklearn 0.22.1" package for the remaining ML algorithms. A bootstrap resampling technique was used to train and validate the classification models. Participants were randomly allocated into training and testing sets at a 7:3 ratio. The performance of each model was evaluated and compared using the validation dataset. The area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 scores were employed to assess the mode’s ability to predict psychological distress risk. Model calibration was further evaluated by visualizing the predicted versus observed incidence of significant psychological distress.
Model optimization and evaluation
To ensure model generalizability and robustness, we employed 5-fold cross-validation to assess the model’s predictive performance. The entire dataset was randomly partitioned into five equally sized folds. During each fold, four folds were used to train the model, while the remaining fold served as the validation set. This process was repeated five times, ensuring all data points were utilized for both training and validation. To evaluate model discrimination, we implemented ROC analysis and quantified performance using the AUC. Calibration plots were further generated to assess the agreement between predicted probabilities and observed outcomes. Decision curve analysis (DCA) was employed to estimate the model’s clinical utility and net benefit in a real-world setting. Feature importance was analyzed using SHAP, where higher absolute SHAP values indicated features with a greater influence on the model’s prediction scores(Lundberg et al., 2020). Additionally, we explored the distribution of feature values and their correlations with model predictions to gain further insights into model behavior.
Statistical analyses
Continuous variables were represented as median (inter-quartile range) and analyzed using the Mann-Whitney U test because all these data did not follow normal distribution. Categorical variables were expressed as number (percentage) and analyzed using the chi-squared test or Fisher’s exact test, as appropriate. All statistical analyses were performed using SPSS 26.0 (IBM Corp., Armonk, NY, USA). A two-sided p-value of < 0.05 indicated statistical significance.