Cancer-related Dysfunctional Beliefs and Attitude about Sleep-6 (C-DBAS-6): a practical and accurate shortened version using XGBoost and SymScore

Abstract

We aimed to develop a practical, data-driven, shortened version of the Cancer-related Dysfunctional Beliefs and Attitudes about Sleep (C-DBAS) scale that maintains diagnostic accuracy while minimizing assessment time and efforts for both patients and clinicians. A sample dataset of 564 cancer patients was collected. Responses to 18 items were organized into six groups based on response similarity using exploratory factor analysis and K-means clustering. The most representative item from each group was then selected utilizing eXtreme Gradient Boosting (XGBoost). Subsequently, a symbolic regression-based clinical score generator (SymScore), a newly developed clinical score generator, was employed to assign optimized weights to the selected items, enabling accurate prediction of the total scores for the Dysfunctional Beliefs and Attitudes about Sleep-16 items (DBAS-16) and 2-item Cancer-related Dysfunctional Beliefs about Sleep (C-DBS) questionnaires. Six key items (items 4, 5, 7, 9, and 15 from DBAS-16 and item C2 from C-DBS) were identified, allowing close estimation of the total score for the combined DBAS-16 and C-DBS, referred to as C-DBAS-6. XGBoost applied to C-DBAS-6 demonstrated strong predictive performance, achieving an R² value of 0.88 when contrasted with the actual total scores of the combined DBAS-16 and C-DBS. Application of the SymScore to C-DBAS-6 achieved comparable performance with an R² value of 0.90, despite being a simpler approach that necessitated only the summation of response weights from a score table without employing a complex machine learning algorithm. The application of SymScore to C-DBAS-6 represents a highly accurate shortened version of the DBAS-16 and C-DBS questionnaires.

Introduction

Insomnia is a common psychiatric symptom reported by patients with cancer [1]. Throughout the trajectory of cancer diagnosis, treatment, and recovery, insomnia can manifest at any stage [2, 3]. Depression or anxiety, observable during diagnosis, can impair sleep, whereas fatigue resulting from chemotherapy or surgery further worsens sleep disturbance [4]. Implementation of the Cognitive–Behavioral Therapy for Insomnia (CBT-I)—a structured, evidence-based approach to treating insomnia without medication—is a crucial and effective strategy for managing sleep disorders [5]. Particularly, the cognitive therapy component of CBT-I can be applied to restructure maladaptive thoughts or dysfunctional beliefs about sleep that hinder sleep maintenance. Given that concerns regarding sleeplessness can influence psychological well-being [6, 7], it is crucial to explore sleep-related maladaptive thoughts that viciously disrupt sleep.

Patients with cancer may experience sleep-related cognition associated with insomnia. The Dysfunctional Beliefs and Attitudes about Sleep-16 (DBAS-16) scale is commonly utilized to assess individuals’ sleep-related cognition [8], and it can be used also to assess cancer patients’ dysfunctional beliefs about sleep [9]. However, sleep-related cognition observed among patients with cancer differs from that among those with insomnia in general [10]. Patients may harbor worries about disease progression, which can affect their perception of sleep disturbances. The Cancer-related Dysfunctional Beliefs about Sleep scale (C-DBS) was developed to assess cancer-specific dysfunctional beliefs regarding sleep and includes two items related to sleep-related concerns: item 1 pertains to immune dysfunction (“My immune system will have serious problems if I don’t go to sleep at a certain time”) and item 2 addresses cancer recurrence (“If I don’t sleep well at night, my cancer may recur or metastasize”). We have previously reported the 14-item composite scale using the DBAS-16 and C-DBS, referred to as the Cancer-related Dysfunctional Beliefs and Attitudes about Sleep-14 (C-DBAS-14) scale [11], with two items of the C-DBS clustering as a separate factor. Patients with cancer, especially those suffering from fatigue due to their disease and chemotherapy, may struggle with hefty assessments involving extensive rating scales.

The development of these shortened questionnaires assessing dysfunctional beliefs about sleep involved different approaches. The items in the DBAS-16 were selected based on several psychometric criteria, clinical relevance, and practical utility [8]. However, the length of DBAS-16 can be burdensome in clinical practice. To address this, a shortened version, DBAS-6, was developed using Exploratory Factor Analysis (EFA) and XGBoost, resulting in high accuracy in assessing dysfunctional beliefs about sleep with fewer items [12]. To assess cancer-related dysfunctional beliefs about sleep, the C-DBAS-14 was developed by integrating the DBAS-16 and C-DBS, and then shortening it to 14 items using psychometric criteria and clinical relevance [11]. However, despite this reduction, the C-DBAS-14 is still relatively lengthy. In this study, we aimed to develop a shorter version using the same approach as the DBAS-6 which has been proven to closely approximate the original scale. We take one step further by applying a recently developed algorithm to create a transparent questionnaire for assessing cancer-related dysfunctional beliefs about sleep for use in clinical settings without sacrificing accuracy. To this end, machine-learning-based prediction models have been proposed [12–15] demonstrating high accuracy in predicting total scores by effectively accounting for the varying importance of individual items. However, these models often rely on specialized algorithms requiring machine learning expertise and additional resources for implementation, which leads to increased labor and financial costs. Moreover, their “black-box” nature limits transparency and interpretability, hindering their adoption in clinical practice.

A symbolic regression-based clinical score generator (SymScore) has been recently proposed to create a shortened version with a transparent algorithmic structure [16]. In contrast to the black-box nature of traditional machine learning, symbolic regression utilizes mathematical operations and user-defined constraints to create interpretable and transparent models. SymScore leverages symbolic regression to assign optimized weights to responses to a shortened questionnaire, thereby predicting the actual total score of the original questionnaire. This approach culminates in a score table from which clinicians can estimate the total score of the original questionnaire by merely adding the assigned weights to the responses from the shortened questionnaire. Thus, the score table generated from SymScore can be efficiently used at the bedside, eliminating the need for complex computing equipment, reducing time and effort required from both clinicians and patients, and enabling rapid diagnosis.

The objective of this study is to use SymScore to develop a transparent and easy-to-use, shortened version of the combined DBAS-16 and C-DBS that effectively captures key aspects of dysfunctional beliefs regarding sleep in patients with cancer.

Methods

Participants and procedure

This retrospective medical records review study was conducted among patients with cancer attending the Sleep Clinic at ASAN Medical Center, Seoul, from January 1, 2017 to September 30, 2024. Inclusion criteria were as follows: 1) patients with cancer aged 18–79 years and 2) those who completed the rating scales. Exclusion criteria were as follows: 1) patients who were unable to move independently, 2) those who had severe medical conditions or organic brain disorders that may impair cognitive function, 3) those with major psychosis or delirium, 4) those unable to complete self-rating scales, or 5) those who could not communicate effectively. A total of 564 patients with cancer met these criteria and had their medical records reviewed. Demographic data were collected, including age, sex, cancer type, cancer stage, current treatment modalities, surgical history, and responses to rating scales (Table 1). The study protocol was approved by the Asan Medical Center Institutional Review Board (IRB, approval no. 2024-1176). Obtaining written informed consent was waived by the IRB, since it is a retrospective medical records review study. This study was conducted in accordance with the tenets of the Declaration of Helsinki.

Table 1.

Demographic and clinical characteristics of the participants (n = 564)

Variable	N (%), mean ± SD
Female sex	399 (70.7%)
Age (years)	54.9 ± 12.1
Cancer types
Solid tumor	528 (93.6%)
Breast cancer	256 (48.5%)
Hepato-biliary cancer	72 (13.6%)
Pulmonary cancer	55 (10.4%)
Gastro-esophageal cancer	47 (8.9%)
Intestinal cancer	33 (6.3%)
Gynecologic cancer	21 (4.0%)
Head and neck cancer	10 (1.9%)
Renal cancer	9 (1.7%)
Prostate cancer	7 (1.3%)
Bladder cancer	6 (1.1%)
Thyroid cancer	5 (1.0%)
Others	7 (1.3%)
Hematologic malignancy	36 (6.4%)
Cancer stages (among 443 patients with available information)
Stage 0	26 (5.9%)
Stage I, II, III	312 (70.4%)
Stage IV	105 (23.7%)
Surgery within 3 months	265 (47.0%)
Current cancer treatment, presence
Chemotherapy	273 (48.4%)
Radiation therapy	88 (15.6%)
Immunotherapy	41 (7.3%)
Anti-hormonal therapy	140 (24.8%)
Psychiatric diagnosis
Insomnia or sleep disorder	283 (50.2%)
Major depressive disorder	143 (25.4%)
Anxiety disorder	46 (8.2%)
Adjustment disorder or somatic symptom disorder	45 (8.0%)
Other	2 (0.4%)
No psychiatric diagnosis	45 (8.0%)
Questionnaires, score
Combined DBAS-16 and C-DBS	95.9 ± 32.0 (20 ~ 176)
C-DBAS-14	74.2 ± 24.8 (5 ~ 138)
ISI	15.5 ± 5.9 (0 ~ 28)

DBAS-16 dysfunctional beliefs and attitude about sleep-16 items, C-DBS cancer-related dysfunctional beliefs about sleep, C-DBAS-14 cancer-related dysfunctional beliefs and attitude about sleep-14 items, ISI insomnia severity index

Measures

DBAS-16

The DBAS-16 is a self-report rating scale designed to assess an individual’s dysfunctional beliefs about sleep [8] (Table 2). It contains 16 items, scored from 0 (strongly disagree) to 10 (strongly agree). A higher final score, representing the average score of all 16 items, reflects a higher level of sleep-related cognition. The validated Korean version of the scale was used in this study [17].

Table 2.

Dysfunctional beliefs and attitudes about sleep-16 (DBAS-16), cancer-related dysfunctional beliefs about sleep (C-DBS), C-DBAS-14 items, and C-DBAS-6 items, and exploratory factor analysis (EFA) results

Scale	Item	Factor						Group	C-DBAS-14	C-DBAS-6
		1	2	3	4	5	6
DBAS-16	1. Need 8 h of sleep	0.59	0.13	-0.00	0.07	0.10	0.02	G1	✓
	2. Need to catch up on sleep loss	0.78	0.04	0.04	0.02	0.13	-0.06		✓
	5. Insomnia interferes with daytime functioning	0.39	0.25	0.23	0.07	0.19	0.08			✓
	3. Consequences of insomnia on health	0.29	0.62	0.07	0.11	0.20	0.15	G2	✓
	4. Fear of losing control over sleep	0.19	0.72	0.56	0.04	0.30	0.04		✓	✓
	6. Better taking sleeping pills	0.10	0.28	0.56	0.26	0.23	0.08	G3	✓
	15. Medication as a solution	-0.02	0.12	0.73	-0.01	0.44	0.07		✓	✓
	7. Mood disturbances due to insomnia	0.22	0.17	0.23	0.65	0.49	0.06	G4	✓	✓
	8. One poor night disturbs whole week	0.30	0.23	0.15	0.22	0.59	0.10	G5	✓
	9. Cannot function without a good night	0.29	0.09	0.13	0.13	0.69	0.09		✓	✓
	10. Sleep is unpredictable	-0.02	0.34	0.14	0.09	0.48	0.07
	11. Unable to manage consequences	0.08	0.31	0.09	0.02	0.76	0.14
	12. Lack of energy due to poor sleep	0.20	0.21	0.08	0.32	0.60	0.19		✓
	13. Insomnia resulting from chemical imbalance	0.06	0.11	0.30	0.06	0.45	-0.02		✓
	14. Insomnia destroying life	0.15	0.21	0.25	0.03	0.71	0.10		✓
	16. Cancel obligations	0.15	-0.00	0.25	0.10	0.57	0.13
C-DBS	Cancer 1. Problems with immune system without sleep	0.05	0.04	0.06	0.07	0.16	0.87	G6	✓
	Cancer 2. Cancer may recur or metastasize	-0.02	0.13	0.04	-0.00	0.11	0.87		✓	✓

DBAS-16 dysfunctional beliefs and attitudes about sleep-16 items, C-DBS cancer-related dysfunctional beliefs about sleep, C-DBAS-14 cancer-related dysfunctional beliefs and attitude about sleep-14 items, C-DBAS-6 cancer-related dysfunctional beliefs and attitudes about sleep-6 items

C-DBS

The C-DBS [10] is an ultra-brief two-item scale designed to assess cancer-specific dysfunctional beliefs regarding sleep in patients with cancer (Table 2). It contains two items: item C1—“My immune system will have serious problems if I don’t go to sleep at a certain time” (immune dysfunction), and item C2—“If I don’t sleep well at night, my cancer may recur or metastasize” (cancer recurrence). The average score of these two items reflects a higher level of sleep-related cognition. The original Korean version was applied in this study.

C-DBAS-14

The C-DBAS-14 (Table 2) is a composite rating scale derived from DBAS-16 (items 1, 2, 3, 4, 6, 7, 8, 9, 12, 13, 14, and 15) and C-DBS (items C1 and C2) [11]. We previously developed the C-DBAS-14 scale by selecting 14 items from all 18 items (16 items of the DBAS-16 and 2 items of C-DBS) using factor analysis and excluding 4 items with extremely high scores or low factor loadings. The average score reflects a higher level of sleep-related cognition related to cancer. We applied the Korean version in this study [11].

Insomnia severity index

The Insomnia Severity Index (ISI) is a self-rating scale designed to assess the severity of insomnia [18]. It contains seven items scored on a five-point Likert scale, with a higher total score indicating greater insomnia severity. The total score range 0–7 indicates no clinically significant insomnia, 8–14 indicates subthreshold insomnia, 15–21 moderate insomnia, and 22–28 severe insomnia. The validated Korean version was applied in this study [19].

Statistical analysis and machine learning algorithms

This section describes the development process of a shortened questionnaire for C-DBS: (1) grouping items based on identified sub-factors, (2) selecting key items to minimize redundancy, and (3) establishing a scoring system for clinical utility. Statistical analysis and machine learning algorithms were conducted using SAS v9.4 (SAS Institute Inc., Cary, NC, USA), R v4.3.2 (R Foundation for Statistical Computing, Vienna, Austria), RStudio (Posit Software, PBC formerly RStudio, PBC, Boston, MA, USA) and Python v3.11.3 (Python Software Foundation, 9450 SW Gemini Dr, Beaverton, Oregon 97008, USA.).

Item grouping based on exploratory factor analysis and K-means clustering

The first step in developing the shortened questionnaire was to identify the sub-factors of the original scales (DBAS-16 and C-DBS) to adequately reflect the sub-domain items of the original scales, and group similar items. For this, we assessed the similarity of response patterns among the 18 items (16 items from DBAS-16 and two items from C-DBS: C1 and C2) using exploratory factor analysis (EFA) (Fig. 1b, heat map). EFA condensed the data into a smaller set of latent factors that could explain the observed correlations among items without predefined assumptions about factor structure. Factors were extracted using principal component analysis (PCA), widely used for its computational efficiency, permitting identification of the minimum number of factors required to explain > 90% of the response variance. Each item was subsequently grouped according to its factor loadings on each factor. The similarity of factor loadings in the identified factor structure represented each item as a point within a multi-dimensional vector space.

Fig. 1.

Key item selection for the shortened questionnaire for DBAS-16 and C-DBS. a Combined DBAS-16 and C-DBS questionnaire assesses dysfunctional beliefs about sleep specific to patients with cancer. To create a shorter version, responses were gathered from 564 patients with cancer (right). b Prior to creating a shortened version, the 18 items from the DBAS-16 and C-DBS that shared common features in latent space were grouped. These common features were extracted through Exploratory Factor Analysis with six factors (left). Subsequently, K-means clustering (k = 6) was applied to categorize these items into six distinct groups (right). c To identify a representative item from each group and minimize redundancy, one item per group was selected and used as input for training an XGBRegressor model. This process was repeated for every possible combination, selecting one item per group. The model performance for each combination was evaluated using the coefficient of determination (R² score). The combination with the highest R.² score (0.88), which includes items 4, 5, 7, 9, 15, and C2, was chosen as the optimal set of items for the shortened questionnaire. C-DBS cancer-related dysfunctional beliefs about sleep, DBAS-16 dysfunctional beliefs and attitudes about sleep-16

Upon determining the factors, K-means clustering was applied to group the items into distinct clusters. The K-means algorithm began with the random selection of K centers within the latent space. Each item’s factor loading was then assigned to the nearest center, forming initial clusters. Each center was updated by calculating the average of the factor loadings assigned to it. This process was repeated until the centers converges, resulting in K clusters, where items with similar factor loadings were grouped together.

Machine learning algorithms to select key items in each group of items

The second step in developing the shortened questionnaire involved selecting a representative item from each group with similar response patterns to avoid redundancy. The aim was to identify an optimal set among all possible item combinations that would accurately predict the total score from all 18 items of the DBAS-16 and C-DBS scales. To assess the predictive performance of each item combination, XGBRegressor, a widely used machine learning algorithm for regression, was employed. Models for each item combination were trained using a 7:3 split of the data into training and testing sets, applying fivefold cross-validation to ensure robustness. For each combination, the coefficient of determination ($R^2$) was computed to measure prediction accuracy, with higher $R^2$ values indicating better alignment between predicted and actual scores. Hence, the item combinations were ranked according to their $R^2$ values and the combination achieving the highest $R^2$ value was selected as the optimal set of items for predicting the total score from the original questionnaire (Fig. 1c).

SymScore: a transparent and accurate scoring algorithm

The final step involved the development and validation of a score table for the selected key items, designed for straightforward application in clinical settings. A scoring system was created to estimate the total score of the original DBAS-16 and C-DBAS scales using solely the selected key items. To achieve this, SymScore, a symbolic regression-based algorithm was applied to automatically assign weights to responses based on predictive importance while imposing necessary constraints and fitting models via symbolic regression. SymScore applies processes of mutation, crossover, and selection to determine optimal weights for each response on the 0–10 point Likert scale of the selected key items. This enables the creation of a score table containing these assigned weights, allowing for the quick calculation of a patient’s total score based on their responses. This approach ensures both predictive accuracy and user-friendliness while providing a transparent alternative to black-box machine learning models. It does not require extensive computational resources or specialized expertise, making it both efficient and easy to use in clinical settings.

Results

Data description

In this study, data was collected from 564 patients with cancer who completed both the DBAS-16 and C-DBS questionnaires to evaluate sleep-related cognition (Table 1 and Fig. 1a). Demographic characteristics of the patients are presented in Table 1. Mean age of participants was 54.9 ± 12.1 years, with 70.7% being female. Regarding cancer type, 93.6% were solid tumors, whereas 6.4% were hematologic malignancies. Approximately half (50.2%) of the participants suffered from insomnia or sleep disorders, followed by major depressive disorder (25.4%), anxiety disorder (8.2%), and adjustment disorder or somatic symptom disorder (9.0%).

Development of C-DBAS-6

In the DBAS-16 and C-DBS questionnaires, each question was rated on a Likert scale ranging from 0 (strongly disagree) to 10 (strongly agree). The total score based on the participants’ responses to the combined 18-item questionnaire represents their sleep-related cognition. The principal objective of this study was to develop a shortened questionnaire capable of quickly and accurately predicting this total score. To create the shortened questionnaire, the number of items was reduced by identifying key items that were crucial to the prediction of the total score of the combined DBAS-16 and C-DBS questionnaires. First, EFA was conducted, and six factors were extracted based on eigenvalues greater than 1.0 and a total variance explained of at least 90%. These six factors represent shared characteristics among items, with factor loadings quantifying the relationship of each item to these factors. Next, items with similar factor loadings were clustered by applying the K-means clustering algorithm with $K=6$. This process resulted in the following six groups: Group 1 (items: 1, 2, 5), Group 2 (items: 3, 4), Group 3 (items: 6, 15), Group 4 (item: 7), Group 5 (items 8, 9, 10, 11, 12, 13, 14, 16), and Group 6 (C1, C2) (Table 2 and Fig. 1b). One representative item was selected from each group to form a unique combination of items. All possible item combinations were evaluated for their predictive performance using XGBRegressor, with the R² value serving as the evaluation metric. Specifically, the R² value was calculated to compare the predicted scores from each combination against the actual total scores of the combined DBAS-16 and C-DBS questionnaires. The optimal combination, referred to as C-DBAS-6, included items 4, 5, 7, 9, and 15 from DBAS-16 and item C2 from C-DBS (Fig. 1c).

The shortened version of the C-DBAS-6 showed a good reliability with an internal consistency of Cronbach’s alpha = 0.740.

Performance of XGBoost on C-DBAS-6

Applying eXtreme Gradient Boosting (XGBoost) to C-DBAS-6 yielded an R² value of 0.88 on the testing set and explained > 90% of the variance in responses. A comparison of the predicted values with the actual total scores from the combined DBAS-16 and C-DBS questionnaires was conducted through a scatter plot (Fig. 2b). The plotted points were tightly clustered around the diagonal, demonstrating high predictive accuracy. Furthermore, a comparison of XGBoost’s performance on C-DBAS-6 with that on C-DBAS-14, a previously shortened version of the original questionnaire, was performed. Notably, while the performance of XGBoost on C-DBAS-6 achieved an R² value of 0.88, the R² value for C-DBAS-14 was recorded at 0.95 (Fig. 2a). Remarkably, XGBoost sustained strong performance with C-DBAS-6, despite using less than half the number of items as C-DBAS-14.

Fig. 2.

Predictive performance of XGBoost applied to C-DBAS-14, XGBoost applied to C-DBAS-6, and SymScore applied to C-DBAS-6 on the testing set. Scatterplots display the relationship between the true total scores from the original combined questionnaire (DBAS-16 and C-DBS) and the predicted total scores from each of the shortened questionnaires. XGBoost applied to C-DBAS-14 demonstrates excellent predictive accuracy (R² = 0.95) with 14 items. In contrast, using only six items, C-DBAS-6 maintains strong predictive accuracy with XGBoost (R² = 0.88) and SymScore (R² = 0.90). C-DBAS-14 cancer-related dysfunctional beliefs and attitude about sleep-14 items, C-DBAS-6 cancer-related dysfunctional beliefs and attitudes about sleep-6 items

Performance of SymScore on C-DBAS-6

Despite its high accuracy, the practical application of XGBoost is limited owing to its black-box nature, reducing interpretability as well as necessitating machine learning expertise for model implementation. To address these challenges, a shortened questionnaire was generated using SymScore, which produces a simple scoring table with assigned weights for each item response (Table 4). By summing the weights of participants’ responses to the C-DBAS-6 using this table, an efficient estimate of the total score for the original 18-item combined questionnaire can be obtained. For example, if a patient’s ratings from the C-DBAS-6 items 4, 5, 7, 9, 15, and C2 are 5, 7, 3, 9, 5, and 10, respectively, the corresponding weights of these responses are 15, 17, 13, 34, 15, and 23, respectively, using the score table. The sum of these weights, 117, serves as a prediction for the total score of the combined DBAS-16 and C-DBS questionnaires.

Table 4.

Score table for the shortened version of DBAS-16 and C-DBS derived by SymScore

Response to item	SymScore weight of each item
	Item 4	Item 5	Item 7	Item 9	Item 15	C2
0	3	3	4	3	2	2
1	5	5	7	4	4	4
2	6	5	10	12	7	5
3	8	6	13	15	10	8
4	12	7	13	17	12	11
5	15	12	17	19	15	11
6	18	13	18	23	17	15
7	20	17	24	26	17	15
8	24	23	26	29	23	19
9	25	29	29	34	27	22
10	27	32	32	37	29	23

By adding the weights corresponding to a participant’s responses, we can quickly and accurately estimate the total score of the original questionnaire

C-DBS cancer-related dysfunctional beliefs about sleep, DBAS-16 dysfunctional beliefs and attitudes about sleep-16 items

Figure 3 presents boxplots of the score distributions computed from SymScore applied to C-DBAS-6 and the combined DBAS-16 and C-DBS scales. The boxplots demonstrate that the score range of SymScore applied to C-DBAS-6 (17–180) closely aligns with that of the original scale (20–176). Similarly, the mean of the scores from SymScore applied to C-DBAS-6 (96.9 ± 33.0) is a near estimate of the mean of the original scale (95.9 ± 32.0), emphasizing the strong similarity between their distributions.

Fig. 3.

Distribution of total scores from the combined DBAS-16 and C-DBAS scales, and scores from SymScore applied to the C-DBAS-6. The boxplots illustrate the distributions of scores for both the combined DBAS-16 and C-DBAS scales, and for SymScore applied to C-DBAS-6. The range of total scores of the combined scale (20–176) is closely approximated by the range of total scores from SymScore (17–180). In addition, the mean and standard deviation (SD) for the combined scale (95.9 ± 32.0) and SymScore (96.9 ± 33.0) exhibit minimal differences. C-DBS cancer-related dysfunctional beliefs about sleep, DBAS-16 dysfunctional beliefs and attitudes about sleep-16

SymScore applied to C-DBAS-6 was expected to strongly correlate with the combined DBAS-16 and C-DBS scales, as well as with the composite scale, C-DBAS-14, due to their shared common cognitive constructs of dysfunctional beliefs in cancer patients. The observed high correlations (r = 0.95, p < 0.001 and r = 0.93, p < 0.001, respectively) (Table 3) support this expectation, demonstrating good convergent validity with existing rating scales. Furthermore, SymScore applied to C-DBAS-6 showed a moderate correlation with ISI (r = 0.48, p < 0.001), comparable to the correlation between ISI and C-DBAS-14 (r = 0.50, p < 0.001) and the combined DBAS-16 and C-DBS scale (r = 0.51, p < 0.001). While maladaptive sleep beliefs, as measured by the combined DBAS-16 and C-DBS scale, can contribute to insomnia severity, as measured by ISI, they do not fully determine insomnia. Thus, the moderate correlations between ISI and dysfunctional sleep belief scales were expected. Overall, the computed correlations confirm that SymScore applied to C-DBAS-6 effectively retains the core features of DBAS-16 and C-DBS while maintaining the expected relationship with insomnia severity using fewer items.

Table 3.

Pearson’s correlation coefficients of rating scales in participants

Variables	C-DBAS-6 SymScore	C-DBAS-14	DBAS-16 and C-DBS
C-DBAS-6 SymScore
C-DBAS-14	0.93*
DBAS-16 and C-DBS	0.95*	0.98*
ISI	0.48*	0.50*	0.52*

C-DBAS-6 cancer-related dysfunctional beliefs and attitude about sleep-6 items, DBAS-16 dysfunctional beliefs and attitudes about sleep-16 items, C-DBS cancer-related dysfunctional beliefs about sleep, C-DBAS-14 cancer-related dysfunctional beliefs and attitude about sleep-14 items, ISI insomnia severity index

*p < 0.01

To evaluate the prediction performance of the shortened questionnaire from SymScore, the predicted and actual values on the testing set were compared using a scatter plot (Fig. 2c). The plotted points were tightly clustered around the diagonal line, demonstrating high accuracy, with an R² value of 0.90. This performance slightly surpassed that of XGBoost applied to C-DBAS-6 (Fig. 2b), which achieved an R² value of 0.88. Thus, despite the simplicity of the score table generated by SymScore, its accuracy remains comparable to that of XGBoost, demonstrating that its ease of use does not compromise its predictive power. Furthermore, after shuffling the training and testing sets 10,000 times, both XGBoost and SymScore applied to C-DBAS-6 yielded mean R² values of 0.88 and 0.90, respectively (Fig. 4). Although XGBoost applied to C-DBAS-14 achieved a higher mean R² value of 0.95 (Fig. 4), the performance of C-DBAS-6 remains remarkable, considering it employs fewer than half the number of items as C-DBAS-14. In addition, SymScore applied to C-DBAS-6 demonstrated greater consistency, as indicated by its lowest variance in R² values across different training–testing dataset shuffling.

Fig. 4.

Robustness comparison of XGBoost applied to C-DBAS-14, XGBoost applied to C-DBAS-6, and SymScore applied to C-DBAS-6 across 10,000 different training-test data set shuffling. Violin plots display the R² score distributions for each shortened questionnaire, capturing performance variability across different training-test data set shuffling. XGBoost applied to C-DBAS-14 achieves the highest mean R² score of 0.95 using 14 items. With only six items, C-DBAS-6 shows comparable mean performance with both XGBoost (R² = 0.88) and SymScore (R² = 0.90). Notably, SymScore shows the least variance in R² scores among the three questionnaires. C-DBAS-14 cancer-related dysfunctional beliefs and attitude about sleep-14 items, C-DBAS-6 cancer-related dysfunctional beliefs and attitudes about sleep-6 items

Discussion

In this study, we developed the C-DBAS-6 scale, a shortened six-item version of the combined DBAS-16 and C-DBS questionnaires. The items for C-DBAS-6 were determined using a combination of EFA, K-means clustering, and XGBRegressor. The final scale consists of five items from the DBAS-16 (items 4, 5, 7, 9, and 15) and one item from the C-DBS (item C2). We observed that the C-DBAS-6 has strong predictive performance, achieving an R² value of 0.88 when compared with the actual total scores of the combined DBAS-16 and C-DBS. The application of SymScore to C-DBAS-6 achieved comparable performance with an R² value of 0.90.

Among the 16 items of the DBAS-16, items 4, 5, 7, 9, and 15 were selected in this study. Through EFA, the items of the DBAS-16 and C-DBS were categorized into six groups (Table 2). Previously, we reported a shortened version of the DBAS-16, namely, DBAS-6, which was developed using EFA and XGBoost. This scale includes items 4, 5, 7, 11, 13, and 15 [12]. Notably, items 4, 5, 7, and 15 were included in both C-DBAS-6 and DBAS-6. From the two items of C-DBS, C2 (cancer recurrence) was chosen as one of the final items for C-DBAS-6, owing to its relevance in patients with cancer who frequently express concerns that sleep disturbances may adversely impact their immune function or increase the risk of cancer recurrence. Existing literature supports the notion that sleep disturbances may indeed impair the immune function among patients with cancer [20, 21] and are associated with cancer recurrence or aggressiveness [22]. However, these are dysfunctional beliefs that contribute to insomnia due to negative, unrealistic worries, biases, and faulty sleep-related cognitions [23]. As no existing scale adequately addresses cancer-specific dysfunctional beliefs about sleep, we previously developed a two-item C-DBS scale to measure these beliefs [10]. Among its two items, C1 (immune dysfunction) and C2 (cancer recurrence), we found no significant difference in the R² values when using either C1 or C2 to predict the total score. However, item C2 demonstrated higher relevance due to its stronger specificity to patients with cancer, leading to its selection as the final item included in the C-DBAS-6.

The application of XGBoost to the C-DBAS-6 scale achieved a performance comparable to that of XGBoost applied to a prior shortened version, C-DBAS-14. This is an outstanding accomplishment given that C-DBAS-6 uses fewer than half the items of C-DBAS-14. Despite this strong performance, the black-box nature of XGBoost complicates the interpretation and explanation of its predictions. To circumvent this limitation, we developed a more user-friendly tool for the C-DBAS-6 using the SymScore algorithm [16]. SymScore generated a simple score table for the C-DBAS-6, enabling a straightforward estimation of the total score from the original DBAS-16 and C-DBS scales through simple summation. Notably, the performance of the score table closely matches and even slightly exceeds the performance of XGBoost applied to the C-DBAS-6, demonstrating that its simplicity does not sacrifice accuracy. By assigning weights to each response and aggregating them, the scoring method becomes not only easier to implement but also more transparent compared to machine learning models. The assigned weights clearly indicate each item’s contribution to the total score. For example, the score table for C-DBAS-6 (Table 4) demonstrates that item 9, when rated with a value of 10, makes the largest contribution to the estimated total score among all items. This intuitive scoring system significantly enhances the interpretability and clinical utility of the C-DBAS-6 scale. Moreover, this approach is both faster and more efficient compared to using the original DBAS-16 and C-DBS questionnaires while also being easier to implement than machine-learning-based shortened versions. The advantages of this approach allow its direct application at the patient’s bedside without the need for computing devices, thereby reducing time requirements while maintaining accuracy.

While our framework is specifically applied to cancer-related dysfunctional beliefs, it holds broader implications for streamlining questionnaire-based assessments in clinical settings. Similar approaches have been successfully employed to condense key questionnaires for other sleep disorders, including tools for evaluating the risk of obstructive sleep apnea, primary insomnia, and comorbid insomnia and sleep apnea [16]. These prior applications highlight the potential of our methodology beyond cancer-related insomnia. Moreover, since our framework is applicable to any quantitative questionnaire-based assessment, it can be extended to other medical conditions. For instance, severity classification of COVID-19 patients has been proposed for efficient triage which requires patients to complete at least 20 questions [24]. The structure of this questionnaire is similar to the questionnaire used in this study. Thus, we expect that our method could optimize such assessments by reducing the burden on patients while maintaining diagnostic accuracy.

While these results are promising, some limitations must be acknowledged that could impact the generalizability of the findings. The sample exclusively comprised Korean individuals, which confines the study to a specific cultural and healthcare context. Sleep-related beliefs among patients with cancer can be influenced by cultural, environmental, and societal factors, meaning that the scoring table developed here may not be directly applicable to other cultural contexts. Future research should explore the utility and validity of C-DBAS-6 within diverse populations to assess its broader generalizability. Furthermore, although the C-DBAS-6 demonstrated strong performance within this data set, additional validation using external data sets is needed to further substantiate its reliability.

In conclusion, the shortened version C-DBAS-6 is well-developed to provide a quick and accurate approximation of the total score derived from the DBAS-16 and C-DBS. This development enables clinicians to efficiently assess cancer-related sleep concerns while minimizing the patient burden. Furthermore, the SymScore approach employed for C-DBAS-6 effectively predicts sleep-related cognition among cancer patients with high accuracy and consistency. In addition, its simplicity and transparency render it well-suited for seamless integration into clinical settings.

Author contributions

Conceptualization: Olive R. Cawiding, Hyeontae Jo, Saebom Jeon, Faeyza Rishad Ardi, Jae Kyoung Kim, Seockhoon Chung; data curation: Seockhoon Chung; formal analysis: Olive R. Cawiding, Hyeontae Jo, Saebom Jeon, Faeyza Rishad Ardi, Jae Kyoung Kim; methodology: Olive R. Cawiding, Hyeontae Jo, Saebom Jeon, Faeyza Rishad Ardi, Jae Kyoung Kim, Seockhoon Chung; funding: Jae Kyoung Kim. Writing—original draft: all authors. Writing—review and editing: all authors.

Funding

J.K.K. is supported by the Institute for Basic Science IBS-R029-C3 and the National Research Foundation of Korea (NRF) (RS-2022-NR068758). H.J. is supported by the National Research Foundation of Korea RS-2024-00357912 and Korea University grants [Grant Numbers.: K2418321 and K2425881]. S.B.J. is supported by the National Research Foundation of Korea [Grant Number: 2022R1F1A1065520]. O.R.C. is supported by the Hyundai Motor Chung Mong-Koo foundation.

Declarations

Conflict of interest

The authors have no competing interests to declare.

Ethical approval

The study protocol was approved by the Asan Medical Center Institutional Review Board (IRB, approval no. 2024-1176). Obtaining written informed consent was waived by the IRB, since it is a retrospective medical records review study.