INTRODUCTION
Insomnia is a common and burdensome symptom experienced by cancer patients during both diagnosis and treatment [1]. In Korea, the prevalence of insomnia among the 10 most common cancer types ranges from 5.8% to 15.2%, depending on cancer type, treatment modality, and the diagnostic criteria used [2]. Despite its clinical significance, insomnia has often been overlooked compared to other psychiatric conditions such as major depressive disorder or anxiety. Because psychological symptoms are typically assessed using subjective rating scales, cancer patients are often required to complete multiple lengthy questionnaires, adding to their burden during an already distressing time. These patients frequently experience psychological distress during diagnosis and treatment [3]. For effective psychological care, it is important to consider a broad spectrum of factors, including psychiatric symptoms, physical complaints, environmental factors, family support, quality of life, or individual coping styles [4]. However, the volume and complexity of conventional rating scales can overwhelm patients when administered all at once [5]. In this context, developing and validating shortened versions of key psychological scales, while preserving diagnostic accuracy, can greatly enhance their usability in clinical settings.
One widely used tool for assessing insomnia is the Insomnia Severity Index (ISI), a 7-item self-report scale that evaluates multiple aspects of insomnia such as sleep onset (item 1a), maintenance of sleep (item 1b), early awakening (item 1c), satisfaction with sleep (item 2), interference with daily function (item 3), other people’s perceptions of the impairment (item 4), and concerns about sleep problems (item 5) [6]. A validated Korean version of the ISI exists for the general population [7]. Moreover, its utility has been demonstrated in various clinical populations, including cancer patients in other countries [8].
To further reduce patient burden and support digital implementation, several shortened ISI versions have been proposed (Table 1) [6,9–11]. Kraepelien et al. [9] selected items 2 and 3; Morin et al. [6] suggested items 1a, 1b, and 1c; Thakral et al. [10] proposed items 2, 3, and 5 in an older population with osteoarthritis; and Wells et al. [11] chose items 2, 4, and 5. More recently, Jo et al. [12] applied machine learning techniques to general population data and developed a 3-item version, ISI-3m, comprising items 1b, 3, and 5, which demonstrated the best predictive performance among the shortened forms. However, its predictive performance has not been evaluated in cancer populations, who may differ significantly in symptom presentation and response patterns. In particular, cancer patients often experience heightened psychological distress, treatment-related fatigue, and disease-specific concerns that may influence how they respond to insomnia-related items [13,14]. Therefore, it remains unclear whether the ISI-3m, originally derived from general population data, can accurately assess insomnia severity in oncology settings.
As part of our validation effort, we first examined the psychometric properties of the full ISI among Korean cancer patients through exploratory and confirmatory factor analyses (EFA and CFA). Factor retention was guided by parallel analysis, model fit was evaluated using standard indices, and internal consistency was assessed with McDonald’s omega. We then assessed the predictive utility of ISI-3m, a machine learning–based shortened version, by training classification models using eXtreme Gradient Boosting (XGBoost), a gradient boosting algorithm known for its high performance and interpretability. In addition to evaluating the reliability and validity of both the full ISI and ISI-3m, we also compared the predictive performance of ISI-3m with other previously proposed abbreviated versions to determine the most effective tool for classifying insomnia severity in cancer care.
METHODS
Participants and Procedure
We conducted a retrospective medical record review of cancer patients aged 18 to 80 years who visited the Stress Clinic for cancer patients in Asan Medical Center. At their initial clinical visit, patients routinely underwent psychological assessments using standardized self-report rating scales. For this study, we collected responses to the ISI from patients who visited the clinic between January 1, 2017 and September 30, 2024. Patients were excluded if they 1) could not move by themselves, 2) had cognitive impairment, 3) were in delirium or psychosis, 4) could not complete self-rating scales, or 5) had communication difficulty. Additionally, more clinical data were obtained including sex, age, marital status, cancer diagnosis, cancer stages, current treatment modalities, and psychiatric diagnoses. The study was approved by the Institutional Review Board (IRB) of Asan Medical Center (2025-0333), and the requirement for written informed consent was waived by the IRB.
Measures
ISI and ISI-3m
The ISI is a self-report questionnaire designed to assess the severity of insomnia symptoms [6,7]. It consists of seven items which can be scored on a five-point Likert scale, yielding a total score ranging from 0 to 28, with higher scores indicating greater insomnia severity. Consistent with prior validation studies, cutoff scores of ISI ≥8 and ≥15 were used to define mild and moderate insomnia, respectively [6]. In this study, we also utilized the ISI-3m, a shortened, machine learning-derived version of the ISI [13]. The ISI-3m was developed using data from the general population through EFA and XGBoost. It comprises three items 1b, 3, and 5, which are identified as the most informative for predicting insomnia severity. This shortened version has demonstrated strong performance in distinguishing individuals with and without insomnia across both binary and multiclass classification settings.
Statistical analysis and machine-learning algorithm
All statistical and machine learning analyses were conducted using Python 3.11.3 (Python Software Foundation) (sklearn 1.5.2, XGBoost 2.1.4, Factor Analyzer 0.5.1) and JASP version 0.14.1.0 (JASP Team).
Reliability and validity of the ISI among cancer patients
Item-level normality was assessed by examining skewness and kurtosis, with values within ±2 considered acceptable. The adequacy of the dataset for factor analysis was verified using the Kaiser–Meyer–Olkin (KMO) measure and Bartlett’s test of sphericity. We conducted EFA to investigate the underlying structure of the ISI. Factor retention was guided by parallel analysis, which compared observed eigenvalues with those generated from simulated random datasets of the same size. The 95th percentile of the simulated distribution was used as the retention threshold. To confirm the structure, CFA was performed using diagonally weighted least squares estimation. Acceptable model fit was defined as standardized root mean square residual (SRMR) ≤0.05, root mean square error of approximation (RMSEA) ≤0.10, and both comparative fit index (CFI) and Tucker–Lewis index (TLI) ≥0.90. Internal consistency was assessed using McDonald’s omega coefficient.
Validation of the performance of ISI-3m
Model description
To assess the predictive performance of the ISI-3m, we train classification models using XGBoost. This model performs through an ensemble of decision trees iteratively, with each tree aiming to correct the errors of its predecessors. This iterative process continually adjusts and improves the model, enabling XGBoost to progressively reduce prediction errors. In addition to strong predictive performance, XGBoost offers insight into the decision-making process by identifying the most influential features, thereby improving both interpretability and model transparency.
Model training and validation
The dataset was divided into training and test sets using a 70:30 split. The training set was used to develop the model through five-fold cross-validation, which helps ensure robustness and prevents overfitting. The final model configuration, determined by cross-validation performance, was evaluated on the test set to estimate generalization error.
Evaluation scores
To evaluate model performance, we used accuracy as the primary metric, assessing the overall correctness of predictions in both binary and multiclass classification settings. In addition to evaluating the ISI-3m, we compared the predictive performance of various other shortened versions of the ISI to determine which model most effectively classifies insomnia severity.
RESULTS
We collected 865 cancer patients and their mean age was 56.1± 12.6 years (Table 2). Among them, 632 (72.7%) were female, and 808 (93.4%) were diagnosed as solid tumors, 409 (47.3%) were having insomnia, and 220 (25.4%) were having depression.
Reliability and Validity of the ISI among Cancer Patients
Before conducting CFA, normality assumption was confirmed based on the skewness and kurtosis of all items (Table 3). The 0.886 of KMO measure and Bartlett’s sphericity (p<0.001) support that the sampling was adequate and data was suitable for analysis. Parallel analysis results identified a single factor model of the ISI with actual eigenvalues (4.61) greater than the simulated 95th percentile (1.13). Further, the CFA shows that the model fit for a single model of the ISI was good (CFI=0.964, TLI=0.942, RMSEA=0.113, SRMR=0.028). Although the RMSEA value (0.113) was slightly higher than the conventional cutoff of 0.10, other indices indicated excellent model fit (CFI= 0.964, SRMR=0.028). Given the sensitivity of RMSEA to sample size and model complexity, the overall model fit was considered acceptable. Reliability of internal consistency was good based on McDonald’s Omega of 0.914.
Suitability of Items of the ISI-3m Using Machine Learning
We assessed the effectiveness of the ISI-3m, originally developed for general population samples, in predicting insomnia severity among cancer patients by comparing its performance with four existing ISI versions proposed by Kraepelien et al. [9], Morin et al. [6], Thakral et al. [10], and Wells et al. [11] (Table 1). To assess ISI-3m performance alongside the other versions, we trained five separate XGBoost models, each utilizing one of the five shortened ISI versions as input features. The dataset was split into 70% for training and 30% for testing (Fig. 1A).
Specifically, we evaluated its ability to distinguish between two clinically relevant thresholds and to classify insomnia severity across multiple categories. The first threshold distinguished individuals with ISI scores below 8 (negative group) from those with scores of 8 or higher (non-negative group), where a score of 8 indicates the onset of mild insomnia (Fig. 1B, left). The second threshold differentiated individuals with scores below 15 (negative or mild insomnia) from those with scores of 15 or higher (moderate or severe insomnia), with a score of 15 marking the clinical cutoff for moderate insomnia (Fig. 1B, middle). We further evaluated the five shortened versions on a multiclass classification task, categorizing insomnia severity into four levels: negative (0–7), mild (8–14), moderate (15–21), and severe (22–28) (Fig. 1B, right).
Based on these thresholds, we next compared the predictive performance across all versions (Fig. 1C). Across all tasks, the ISI-3m demonstrated the strongest predictive performance, achieving the highest binary classification accuracy: 0.935 for detecting mild insomnia (ISI ≥8) and 0.857 for moderate insomnia (ISI ≥15). For the multi-class classification task, ISI-3m achieved the highest overall accuracy of 0.827. Overall, ISI-3m consistently outperformed the other versions across all classification tasks, indicating that the selected items (1b, 3, and 5) are the most effective for predicting varying levels of insomnia severity among cancer patients (Fig. 2).
DISCUSSION
This is the first study to validate the psychometric properties of the ISI among Korean cancer patients. Our findings confirmed a single-factor structure for the ISI in this population, which contrasts with the two-factor structure previously reported in the general Korean population. This result suggests that cancer-specific symptom patterns, such as heightened psychological distress or treatment-related sleep disruption, may influence how patients interpret insomnia-related items.
Building on these findings, we evaluated the ISI-3m, a machine learning–derived, shortened version of the ISI comprising items 1b, 3, and 5. Originally developed using general population data, the ISI-3m demonstrated strong predictive accuracy for total ISI scores and significantly reduced response burden in our cancer patient sample. This is particularly valuable in oncology settings, where patients often experience elevated fatigue and symptom burden. By retaining the most informative items, the ISI-3m preserves core diagnostic content while improving clinical usability.
To validate the ISI-3m among cancer patients, we applied a machine learning approach using XGBoost to assess its ability to classify insomnia severity. Classification models were trained and tested on independent subsets of the data, allowing us to evaluate model generalizability. ISI-3m achieved high classification accuracy across both binary thresholds (mild and moderate insomnia) and multiclass settings, supporting its use as a rapid screening tool in oncology care. In direct comparison with four other previously proposed abbreviated ISI versions, those by Kraepelien et al. [9], Morin et al. [6], Thakral et al. [10], and Wells et al. [11], ISI-3m consistently outperformed the alternatives. These findings emphasize the robustness and generalizability of ISI-3m, despite differences in item selection across studies. Such variation likely reflects differences in sample demographics, cultural context, and methodological approach.
Machine learning methods have increasingly been employed in clinical survey reduction, offering scalable and data-driven strategies to identify the most informative items without sacrificing validity. Techniques such as gradient boosting and random forests have been successfully used to shorten questionnaires while preserving high classification accuracy and internal consistency [15–17]. These approaches support the clinical utility of abbreviated scales, especially in high-demand environments, by ensuring that reduced forms remain both efficient and psychometrically sound. The consistent success of such models across various insomnia-related measures reinforces the robustness of machine learning–based reduction strategies in optimizing clinical assessments.
This study has several limitations. First, it was conducted in a single tertiary hospital in South Korea, which typically treats patients with more severe conditions; this may limit the generalizability of the findings. Second, as the ISI responses were retrospectively obtained from medical records, potential biases related to self-report and recall cannot be ruled out. Third, all participants were evaluated in a specialized stress clinic, suggesting that psychological symptom severity may be elevated compared to the general cancer patient population. Fourth, the sample was predominantly female, with a high proportion of breast cancer patients. Given evidence that female cancer patients may report greater sleep disturbance and fear of disease progression, these factors should be considered when interpreting the results. Lastly, while machine learning methods such as XGBoost demonstrated strong predictive performance in validating the ISI-3m, their inherent black-box nature limits interpretability and clinical transparency [18]. To address this, future work could explore the use of symbolic regression–based approaches to derive closed-form scoring functions for the ISI-3m. The SymScore framework, which combines machine learning accuracy with human-interpretable equations, has been proposed to resolve the trade-off between performance and transparency [19]. It has also been successfully applied in other clinical survey shortening tasks, supporting its feasibility and potential utility in developing reduced yet interpretable screening tools [20]. By enabling the construction of score tables that do not require computational resources at the point of care, symbolic regression methods could further enhance the usability of the ISI-3m in time-constrained oncology settings and facilitate rapid, transparent insomnia screening.
In summary, this study supports the ISI-3m as a practical, efficient, and psychometrically robust tool for assessing insomnia severity in cancer patients. Its brevity and predictive accuracy make it well-suited for integration into routine oncology care, where quick yet reliable screening tools are essential. These findings extend the utility of the ISI-3m beyond the general population and emphasize its potential as a valuable clinical instrument in supportive cancer care.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation 
Print 
