Evaluation of inflammatory bowel disease-related sleep disorders based on an interpretable machine learning approach: a multicenter study in China

Abstract

Background:

Patients with inflammatory bowel disease (IBD) often encounter complications such as sleep disorders, which are of great detriment to their quality of life, and earlier identification and intervention can effectively improve the prognosis of patients.

Objectives:

In this study, we worked on building a risk model to assess IBD-related sleep disorders using a machine learning (ML) approach.

Design:

Observational study.

Methods:

Based on an online questionnaire, we collected clinical data from 2478 IBD patients from 42 hospitals in 22 Chinese provinces between September 2021 and May 2022. Then, we developed and validated six common ML models to assess the risk of co-morbid sleep disorders in IBD patients, and evaluated and compared the performance of these models using relevant metrics. Finally, the Local Interpretable Model–Agnostic Explanations algorithm (Lime) was utilized to interpret the results of the best ML model.

Results:

In this study, after multidimensional comparisons, the voting model was finally identified as superior among several models, with the area under the curve and accuracy reaching 0.76 and 0.74, respectively. After calculations, it was found that the co-morbidities of depression and anxiety, an older age, outpatient diagnosis, and a longer course of the disease were all indicative of a higher risk of sleep disorders among IBD patients in this model.

Conclusion:

The construction of risk assessment models using ML has high clinical value in the prediction of IBD-related sleep disorders, and the efficacy of its application suggests it can serve as a promising evaluation tool in clinical work.

Plain language summary

Interpretable machine learning analysis of sleep disorders in Chinese patients with inflammatory bowel disease: a multicenter study

Patients with inflammatory bowel disease often suffer from sleep disorders, which significantly impair their quality of life. Early detection and intervention can greatly improve patient outcomes. This study aimed to develop and validate a machine learning model to assess the risk of sleep disorders in IBD patients. Clinical data from 2,478 IBD patients across 42 hospitals in 22 Chinese provinces were collected through an online questionnaire from May 2020 to May 2021. Six common ML models were developed and validated, with the voting model identified as the most effective, achieving an area under the curve of 0.76 and an accuracy of 0.74. Key risk factors for sleep disorders included comorbid depression and anxiety, older age, outpatient diagnosis, and a longer disease duration. The Local Interpretable Model–Agnostic Explanations algorithm was used to interpret the results of the best-performing model. The findings suggest that ML-based risk assessment models have high clinical value for predicting IBD-related sleep disorders and can serve as a promising tool for early evaluation and intervention in clinical practice.

Introduction

Inflammatory bowel disease (IBD) mainly includes ulcerative colitis (UC) and Crohn’s disease (CD) and is a group of chronic nonspecific inflammatory diseases of the intestine in which inflammation persists in the gut. IBD is characterized by intermittent relapses and remissions, commonly occurring in adolescence and adulthood. ¹ The etiology of IBD has not been clarified and can be attributed to the interplay of multiple factors, including individual genetic predisposition, the external environment, the intestinal microbiota, and the immune response. ² After definitive diagnosis, patients often require lifelong treatment, including medications, nutrition, lifestyle changes, and surgical treatment. The goal of treatment is to improve the patient’s quality of life (QOL) by treating exacerbations and preventing remission. ³ IBD is a global disease that was first reported in Western industrialized countries, with a steadily increasing prevalence rate. As the population with the disease ages, there will be a gradual transition toward the equilibrium stage. With the gradual Westernization of developing countries, the prevalence rate has been rising rapidly, although it is still at a low level.^4,5 This will place a heavy and direct economic burden on the healthcare system, as IBD is mostly diagnosed in young adulthood, resulting in indirect costs associated with lost productivity that are likely to equal or even exceed direct healthcare costs. ⁶ Developing countries, where IBD diagnosis and care systems are not well developed, will face even greater challenges.

Sleep disorders are one of the most important parenteral manifestations in patients with IBD, which is characterized by prolonged sleep latency, frequent sleep fragmentation, increased use of sleep medications, and decreased sleep quality overall. ⁷ With the progression and prolongation of IBD, patients’ gastrointestinal (GI) symptoms, such as abdominal pain and increased bowel movements, will lead to an increase in nocturnal awakenings. In addition to the influence of physiological factors, patients’ own perceived stress, the number of rotating nightshift jobs, their mental health, and the frequency of strenuous physical activity are also risk factors for sleep disorders associated with IBD.^8,9 Sleep is a cyclical process by which the body eliminates fatigue, and several studies have shown that sleep can be involved in the regulation of the immune system, either directly or indirectly through the nervous and neuroendocrine systems. When sleep is disrupted, pro-inflammatory signaling increases, immune system homeostasis is disrupted, and some degree of impairment of the gut function ensues. Acute sleep deprivation (overnight sleep deprivation) may cause a transient immune activation in the body,^10,11 while long-term chronic sleep deprivation leads to a decrease in the number of peripheral blood immune cells and changes in cytokine levels, and the longer the period of deprivation, the greater the effect on the immune system,^12,13 thus affecting the course of the patient’s disease. Some studies found that the presence of sleep disorders in patients is mostly associated with disease activity and an increased risk of disease recurrence.^14,15 There is a bidirectional relationship between sleep disorders and IBD, and sleep disorders and inflammation may form a self-perpetuating and vicious cycle that allows the progression of the IBD. Therefore, it is vitally necessary to pay more attention to, and effectively deal with, sleep disorders in IBD patients, which will greatly improve their QOL. Although existing studies have established associations between sleep disorders and factors like depression, disease activity, and inflammation markers, these findings remain largely descriptive and lack translational utility. Crucially, clinicians face two unmet needs: (1) the inability to quantify individualized risk from multiple interacting factors and (2) no standardized protocol to identify high-risk patients warranting sleep interventions.

Artificial intelligence (AI) technologies are now widely used and rapidly developing in the medical field. Among them, machine learning (ML) is one of the important branches of AI, with superiority in recognizing the complex relationship between variables and outcomes. It has been applied in many fields, such as assisted diagnosis, individualized treatment, disease prediction and prevention, drug prevention, medical image analysis, etc. There are also more and more studies confirming that ML models perform well in the context of various diseases.^16,17 However, there are currently no studies related to the application of ML algorithms to assess IBD-related sleep disorders. Therefore, in this multicenter study, we collected relevant data from clinical IBD patients to develop and validate assessment models applicable to IBD-related sleep disorders.

Methods

Data source and study subjects

In this multicenter study, we included IBD patients from 42 hospitals in 22 provinces of China between September 2021 and May 2022, and all participants were surveyed with an online questionnaire. Ethical approval for this study was granted by the institutional review board of the Renmin Hospital of Wuhan University (approval number: WDRY2022-K150), and all the participants signed informed consent forms. This study was also conducted in compliance with the ethical principles of the Declaration of Helsinki. The inclusion criteria of the patients were as follows: (1) IBD was clearly diagnosed according to the consensus approach to the diagnosis and treatment of IBD, which was formulated by the Chinese Society of Gastroenterology in 2018 ¹⁴ ; (2) age ⩾18 years old; (3) all the enrolled patients were able to understand the survey content and agree to the survey, and were willing to receive management from physicians as well as psychological investigations, with good medication compliance; (4) complete clinical data. Further, the exclusion criteria were as follows: (1) previous history of mental illness such as depression and schizophrenia; (2) presence of other diseases that seriously affect the QOL; and (3) undetermined colitis.

Data collection and outcome definition

In this study, an online questionnaire was utilized to collect data related to IBD patients, including demographic characteristics (age, sex), first visit, disease activity or severity (remission, mild activity period, moderate activity period, and severe activity period, which were categorized according to the latest consensus on the diagnosis and treatment of IBD ¹⁸ ), course of disease (less than 2 years, 2–5 years, over 5 years), outpatient diagnosis (UC or CD), symptoms (intestinal symptoms, systemic inflammatory symptoms, extraintestinal symptoms, complication manifestations, depression, anxiety), received treatment (5-aminosalicylic acid (5-ASA), glucocorticoids, immunosuppressant, biological agents, surgery). This retrospective observational study did not include inflammatory markers (e.g., CRP, fecal calprotectin) or sleep medication data. The detection methods and standards for inflammatory markers (e.g., CRP sensitivity, fecal calprotectin sampling timing) may vary significantly across different hospitals. Forcibly incorporating such data could lead the model to learn “detection discrepancies” rather than true biological associations. Similarly, sleep medication use may be a consequence rather than a cause of sleep disturbances, and directly including it could lead the model to conflate true risk factors. This model aims to provide rapid risk assessment in initial diagnosis or primary care settings. In some regions with limited medical resources, immediate access to inflammatory marker testing or detailed medication histories may be unavailable. If the model relies heavily on such features, its generalizability would be compromised. In contrast, the current model, based on demographics, symptoms, and psychological scales, uses features that are easier to standardize and collect, making it more widely applicable.

We systematically assessed depression and anxiety using relevant scales. (1) Patient Health Questionnaire (PHQ-9; Supplemental Table 1): The PHQ-9 is a simple and validated self-assessment scale for depressive disorders based on the fourth edition of the Diagnostic and Statistical Manual of Mental Disorders. The PHQ-9 has good reliability and validity in supporting the diagnosis of depressive disorders and assessing symptom severity.^19,20 The PHQ-9 scale consists of nine questions, each of which can be scored as 0, 1, 2, or 3, with total scores of 0–4, 5–9, 10–14, and 15–27 indicating no depression, mild depression, moderate depression, and severe depression, respectively. It has been shown that major depression rarely occurs below a score of 10, and that a score below 10 can be used as a judgment value for the presence of a significant clinical treatment effect. ²¹ In the present study, we also used a score of 10 as a threshold value for the simple dichotomization of patients. (2) Generalized Anxiety Disorder-7 items (GAD-7; Supplemental Table 2): the GAD-7 is a widely used clinical scale. It can be used to assess patients’ anxiety and to observe changes in condition and treatment outcomes. The assessment scale consists of seven items, each of which can be scored as 0, 1, 2, or 3, with a total score between 0 and 21. A score of 0–4 indicates no anxiety, a score of 5–9 indicates mild anxiety, a score of 10–14 indicates moderate anxiety, and a score of 15–21 indicates severe anxiety. ²² It has been found in several studies that a score of 10 can be considered as a red flag for anxiety disorders.^23,24 Therefore, we take the score of 10 as the threshold in this study. Both the PHQ-9 and GAD-7 have demonstrated reliability, validity, and clinical utility in Chinese populations.^25–27 Moreover, these scales have been widely adopted as objective measurement tools in multiple studies related to IBD.^28,29

In the current study, IBD-related sleep disorders were used as the primary outcome of the ML models, which was assessed by the Pittsburgh Sleep Quality Index (PSQI; Supplemental Table 3) to determine whether patients with IBD suffer from sleep disorders or not. The PSQI scale is widely used clinically to rate the sleep of subjects in the last 1 month, such as those with IBD, Irritable Bowel Syndrome, and functional dyspepsia.^30–32 The scale evaluates patients in seven aspects, including subjective sleep quality, time to sleep, sleep duration, sleep efficiency, sleep disorders, hypnotic drugs, and daytime dysfunction. The total score, which is the cumulative score for each aspect, ranges from 0 to 21, and 0–5, 6–10, 11–15, and 16–21 are categorized as indicating good sleep quality, moderate sleep quality, poor sleep quality, and very poor sleep quality, respectively. ³³ The Chinese version of the PSQI has been rigorously validated in Chinese populations, demonstrating excellent reliability (Cronbach’s α = 0.84) and strong construct validity (with all component factor loadings exceeding 0.5) in psychometric evaluations conducted among Chinese university students. ³⁴ Several previous IBD-related studies have used the PSQI as an objective measure in the assessment of sleep disorders.^35,36 We categorized patients with scale scores between 0 and 6 as having unimpaired sleep quality and more than 6 as having sleep disorders. ³⁴

Data preprocessing and features (variables) selection

The original data involved in this study contained a total of 17 features, including age, sex, first visit, severity, course of disease, outpatient diagnosis, intestinal symptoms, inflammation, extraintestinal symptoms, complication, 5-ASA, glucocorticoids, immunosuppressant, biological agents, surgery, depression and anxiety, and sleep is used as the outcome label. There are no missing values, outliers, extreme values, or data imbalances in all the original data, indicating that the original data have high quality. We preprocessed the categorical features by one-hot encoding, including sex, first visit, outpatient diagnosis, symptoms, treatment, depression, anxiety, and sleep disorders. In order to reduce the influence of dimension and guarantee the comparability, the continuous variable of age was standardized using the Standard Scaler function (x* = $\frac{x-\mu }{\sigma }$ ). Subsequently, the last absolute shrinkage and selection operator (Lasso) analysis was performed on all features to further screen the included features and thus improve the generalization ability and interpretability of the model. Among the 2478 IBD patients included in this study, a total of 1211 patients (48.87%) with sleep disorders were included, and then we randomly assigned the samples into the training dataset (80%) for model establishment purposes and the testing dataset (20%) for evaluating the tuned ML models.

ML model establishment and evaluation

In this study, we used several widely used supervised ML algorithms to construct ML models, including Logistic Regression (LR), K-Nearest Neighbor (KNN), Multilayer Perceptron (MLP), Random Forest (RF), the eXtreme Gradient Boosting (XGBoost), and the soft voting ensemble-based model. LR is a generalized linear regression analysis model, which is mostly used in practice to solve binary classification problems and can be used to construct simple classification models. KNN is one of the simplest methods of data mining classification. It is realized by determining the similarity or distance between the test set and the training set, and it has the advantages of simplicity and clarity, as well as a precise prediction capacity. ³⁷ MLP is a deep learning model based on feed-forward Artificial Neural Network, which consists of multiple neuron layers, in which the input layer accepts the input features, the output layer gives the final prediction, and the intermediate hidden layer is used to extract features and perform nonlinear transformations. MLP has been applied in various areas such as cancer prediction, parallel MRI fast imaging, and synthetic CT generation.^38–41 RF belongs to one of the major branches of ML—integrated learning, the basic unit of which is the decision tree; it builds multiple decision trees and fuses them together to obtain a more accurate and stable model. XGBoost is an ML algorithm based on Gradient Boosting Decision Trees, the core idea of which is to build a powerful prediction model through iteration that is characterized by high efficiency, flexibility, and scalability. In addition, a soft voting model is used in this study, which consists of three models, LR, RF, and MLP. The average probabilities of the three models being able to predict whether a sample has sleep disorders or not are used as criteria.

Grid search, as a means of tuning hyperparameters through multiple cycles, is often combined with cross-validation (CV) to screen the optimal hyperparameter configurations. In this study, grid search combined with 10-fold CV is used to divide the training set into 10 parts, of which 9 parts are used to train the model and the remaining 1 part is used to adjust the parameters so as to obtain the classification performance index (accuracy). After 10 iterations, 10 accuracy values and their averages are obtained. The above process is repeated for all hyperparameter combinations, and the best hyperparameter combinations are filtered and applied to the testing dataset to derive the final ML models.

We here evaluate the performances of the ML models in the testing dataset, and the evaluation metrics used include area under the curve (AUC), Accuracy $\left(\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}\right)$ , Sensitivity (Recall) $\left(\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}\right)$ , Specificity $\left(\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}\right)$ , Precision $\left(\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}\right)$ , and F1 scores $\left(\frac{2\ast \mathrm{Precision}\ast \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}\right)$ , where TP, FP, TN, and FN represent true positives, false positives, true negatives, and false negatives, respectively. For calibration, we plotted calibration curves for all models and computed Brier scores to quantitatively assess the degree of calibration. The Brier score serves as a relative measure that can be used to compare the performances of models  $\left(\mathrm{Brier}\mathrm{score}(Y,P)=1/n{\displaystyle \sum }_{i=1}^n{\left(P_i-Y_i\right)}^2\right)$ . The higher the score, the worse the prediction and the worse the calibration, and so the closer the Brier score is to 0, the better. ⁴²

Feature importance and ML model interpretations

In this study, a total of nine features were selected after screening, and we have used Permutation Importance for feature importance assessment. This method is used to disrupt the relationship between features and labels by randomly disrupting individual features, and observing how much the feature changes affect the accuracy of the mode; this yields a ranking of the importance of the features. To increase the transparency and interpretability of the model, we used the Lime algorithm for the interpretation of individual prediction so as to clarify the process of making classification decisions as undertaken by the ML model in patients with IBD, and to evaluate the reliability of the model. ⁴³

Software and statistical analysis

In the process of ML model construction, data processing, feature selection, statistical analysis, evaluation, and ML model interpretation are performed using Python software, version 3.9.12 (Python Software Foundation). The packages we used included “Pandas,” “NumPy,” “Scikit-learn,” “Matplotlib,” “NumPy,” “Matplotlib,” “SciPy,” etc. The continuous feature was expressed as mean ± standard deviation (mean ± SD), and the values were compared by Student t tests, while the categorical features were expressed with percentages and compared by Pearson’s Chi-square test or Fisher’s exact test. All statistical analyses were two-tailed, with a p value lower than 0.05 deemed statistically significant.

Guideline

The reporting of this study conforms to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement (Supplemental Table 4).

Results

Patient characteristics

In this study, we included a total of 2843 potential patients. Based on the inclusion and exclusion criteria, 285 patients were excluded in the first round of screening, including 154 who did not consent to be surveyed, 71 who were younger than 18 years old, and 60 whose clinical data were incomplete. Subsequently, the medical records of 2558 patients were reviewed for the second round of screening, and a total of 80 patients were excluded, including 11 with a history of psychiatric disorders (8 with depression and 3 with schizophrenia), 8 with diseases seriously affecting the QOL, and 61 with undetermined colitis. Thus, a total of 2478 patients were finally included in this study (1371 cases with UC and 1107 cases with CD). We randomized the patients into two original datasets, the training dataset and the testing dataset, containing 1982 and 496 patients, respectively (Figure 1). There were 1211 patients with sleep disorders out of all included patients, with a mean age of 36 years, of which 40.13% were male, and there were 1267 patients without sleep disorders, with a mean age of 35 years, of which 35.12% were male. The rest baseline characteristics are shown in Table 1. After randomized grouping, there was no statistically significant difference found in the baseline characteristics of patients in the training dataset and validation dataset (Table 2). Among them, 965 (48.69%) patients in the training dataset and 246 (49.60%) in the validation dataset had sleep disorders.

Figure 1.

The process of patient inclusion and exclusion screening.

Table 1.

Baseline characteristics of the enrolled IBD patients.

Characteristics	All patients (n = 2478)	Impaired sleep quality (n = 1211)	Normal sleep quality (n = 1267)	χ2 or Z value	p Value
Age, years	35 (28–46)	36 (29–48)	35 (28–45)	−3.06	<0.01
Male, n (%)	931 (37.57)	486 (40.13)	445 (35.12)	6.41	0.01
First visit, n (%)	493 (19.90)	269 (22.21)	224 (17.68)	7.70	0.01
Severity, n (%)
Remission	946 (38.18)	433 (35.76)	513 (40.49)	12.65	0.01
Mild activity period	588 (23.73)	294 (24.28)	294 (23.20)
Moderate activity period	734 (29.62)	360 (29.73)	374 (29.52)
Severe activity period	210 (8.47)	124 (10.24)	86 (6.79)
Course of disease, n (%)
Less than 2 years	941 (37.97)	423 (34.93)	518 (40.88)	10.51	0.01
2–5 years	764 (30.83)	381 (31.46)	383 (30.23)
Over 5 years	773 (31.19)	407 (33.61)	366 (28.89)
Outpatient diagnosis, n (%)
UC	1371 (55.33)	664 (54.83)	707 (55.80)	0.20	0.66
CD	1107 (44.67)	547 (45.17)	560 (44.20)
Symptom, n (%)
Intestinal symptoms	2113 (85.27)	1048 (86.54)	1065 (84.06)	2.85	0.09
Systemic inflammatory symptoms	108 (4.36)	54 (4.46)	54 (4.26)	0.02	0.89
Extraintestinal symptoms	198 (7.99)	102 (8.42)	96 (7.58)	0.49	0.48
Complication manifestations	205 (8.27)	97 (8.01)	108 (8.52)	0.15	0.70
Depression	737 (29.74)	618 (51.03)	119 (9.39)	511.79	<0.01
Anxiety	631 (25.46)	508 (41.95)	123 (9.71)	337.41	<0.01
Received treatment, n (%)
5-ASA	1531 (61.78)	756 (62.43)	775 (61.17)	0.36	0.55
Glucocorticoids	383 (15.46)	188 (15.52)	195 (15.39)	<0.01	0.97
Immunosuppressant	360 (14.53)	173 (14.29)	187 (14.76)	0.08	0.78
Biological agents	1273 (51.37)	621 (51.28)	652 (51.46)	<0.01	0.96
Surgery	304 (12.27)	160 (13.21)	144 (11.37)	1.79	0.18

5-ASA, 5-aminosalicylic acid; CD, Crohn’s disease; IBD, inflammatory bowel disease; UC, ulcerative colitis.

Table 2.

Baseline characteristics of the enrolled IBD patients in training and testing dataset.

Characteristics	All patients (n = 2478)	Training dataset (n = 1982)	Validation dataset (n = 496)	χ2 or Z value	p Value
Age, years	35 (28–46)	35 (28–46)	36 (28–47)	−0.57	0.57
Male, n (%)	931 (37.57)	729 (36.78)	202 (40.73)	2.47	0.12
First visit, n (%)	493 (19.90)	399 (20.13)	94 (18.95)	0.28	0.60
Sleep disorder, n (%)	1211 (48.87)	965 (48.69)	246 (49.60)	0.10	0.76
Severity, n (%)
Remission	946 (38.18)	759 (38.29)	187 (37.70)	5.75	0.12
Mild activity period	588 (23.73)	452 (22.81)	136 (27.42)
Moderate activity period	734 (29.62)	603 (30.42)	131 (26.41)
Severe activity period	210 (8.47)	168 (8.48)	42 (8.47)
Course of disease, n (%)
Less than 2 years	941 (37.97)	752 (37.94)	189 (38.10)	0.94	0.62
2–5 years	764 (30.83)	619 (31.23)	145 (29.23)
Over 5 years	773 (31.19)	611 (30.83)	162 (32.66)
Outpatient diagnosis, n (%)
UC	1371 (55.33)	1082 (54.59)	289 (58.27)	2.02	0.16
CD	1107 (44.67)	900 (45.41)	207 (41.73)
Symptom, n (%)
Intestinal symptoms	2113 (85.27)	1693 (85.42)	420 (84.68)	0.12	0.73
Systemic inflammatory symptoms	108 (4.36)	85 (4.29)	23 (4.64)	0.05	0.83
Extraintestinal symptoms	198 (7.99)	158 (7.97)	40 (8.06)	<0.01	1.00
Complication manifestations	205 (8.27)	167 (8.43)	38 (7.66)	0.21	0.64
Depression	737 (29.74)	583 (29.41)	154 (31.05)	0.43	0.51
Anxiety	631 (25.46)	490 (24.72)	141 (28.43)	2.68	0.10
Received treatment, n (%)
5-ASA	1531 (61.78)	1237 (62.41)	294 (59.27)	1.52	0.22
Glucocorticoids	383 (15.46)	312 (15.74)	71 (14.31)	0.51	0.47
Immunosuppressant	360 (14.53)	295 (14.88)	65 (13.10)	0.87	0.35
Biological agents	1273 (51.37)	1016 (51.26)	257 (51.81)	0.03	0.86
Surgery	304 (12.27)	246 (12.41)	58 (11.69)	0.13	0.72

5-ASA, 5-aminosalicylic acid; CD, Crohn’s disease; IBD, inflammatory bowel disease; UC, ulcerative colitis.

Feature selection

Here, 17 features were initially included in the study, and after counting all the original data, it was found that the proportion of single samples was less than 10% for the features of inflammation, extraintestinal symptoms, and complication so these three features were removed. Subsequently, the lasso regression analysis was performed on the remaining 14 features, and a total of 9 features were screened out, including age, sex, first visit, severity, course of disease, outpatient diagnosis, surgery, depression, and anxiety (Figure 2).

Figure 2.

The lasso regression analysis of 14 features.

ML model performance

In this study, we used a total of six ML models for the evaluation of sleep disorders associated with IBD, and each model was tuned to obtain the best hyperparameters (Table 3). We compared the performance metrics of all the ML models, as shown in Table 4 and Figure 3. After comparison, voting was found to be the most well-performing model, with the highest AUC, Accuracy, and F1 scores of 0.76, 0.74, and 0.70, respectively. In terms of Sensitivity (recall), Specificity, and Precision, the voting model yielded the highest value. In addition to this, we plotted calibration curves for all models (Figure 4). The Brier score of the voting model was calculated as 0.19, which is tied with LR as the lowest value, indicating that it has the best calibration. We further plotted decision curve analysis (DCA) curves for the voting model, as shown in Figure 5. Within a certain threshold range, the net benefit is higher than both the treat-all and treat-none curves, indicating that the voting model has practical value.

Table 3.

Best hyperparameters of models.

Model	Best hyperparameter
LR	{‘C’: 0.1, ‘max_iter’: 100, ‘penalty’: ‘l2’, ‘solver’: ‘lbfgs’, ‘tol’: 0.0001}
RF	{‘max_features’=2, ‘criterion’: ‘entropy’, ‘n_estimators’: 100}
KNN	{‘algorithm’: ‘brute’, ‘leaf_size’: 30, ‘metric’: ‘minkowski’, ‘n_neighbors’: 3}
XgBoost	{‘colsample_bytree’: 1, ‘learning_rate’: 0.3, ‘max_depth’: 6, ‘n_estimators’: 46, ‘subsample’: 1}
MLP	{‘activation’: ‘relu’, ‘alpha’: 0.001, ‘beta_1’: 0.9, ‘beta_2’: 0.999, ‘epsilon’: 1e-08, ‘max_iter’: 200, ‘solver’: ‘‘lbfgs’, ‘tol’: 0.0001}
LR	C=0.1, class_weight=None, dual= False, fit_intercept=True, intercept_scaling=1, l1_ratio=None, max_iter=100, multi_class=‘auto’, n_jobs=None, penalty=‘l2’, random_state= 0, solver= ‘lbfgs’, tol= 0.0001, verbose= 0, warm_start= False
RF	bootstrap= True, ccp_alpha= 0.0, class_weight= None, criterion=‘entropy’, max_depth= None, max_features=2, max_leaf_nodes= None, max_samples= None, min_impurity_decrease=0.0, min_samples_leaf= 2, min_samples_split= 2, min_weight_fraction_leaf= 0.0, n_estimators= 100, n_jobs= None, oob_score= False, random_state= 0, verbose= 0, warm_start= False
KNN	algorithm= ‘brute’, leaf_size= 30, metric= ‘minkowski’, metric_params= None, n_jobs= None, n_neighbors= 3, p = 2, weights= ‘uniform’
XgBoost	objective= ‘binary:logistic’, use_label_encoder= False, base_score= 0.5, booster= ‘gbtree’, callbacks= None, colsample_bylevel= 1, colsample_bynode= 1, colsample_bytree= 1, early_stopping_rounds= None, enable_categorical= False, eval_metric= None, gamma= 0, gpu_id= −1, grow_policy= ‘depthwise,’ importance_type= None, interaction_constraints= ”, learning_rate= 0.300000012, max_bin= 256, max_cat_to_onehot= 4, max_delta_step= 0, max_depth= 6, max_leaves= 0, min_child_weight= 1, monotone_constraints= ‘()’, n_estimators= 46, n_jobs= 0, num_parallel_tree= 1, predictor= ‘auto’, random_state= 0, reg_alpha= 0, reg_lambda= 1, sampling_method= ‘uniform’, scale_pos_weight= 1, subsample= 1, tree_method= ‘exact’, validate_parameters= 1, verbosity= None
MLP	activation= ‘relu’, alpha= 0.001, batch_size= ‘auto’, beta_1= 0.9, beta_2= 0.999, early_stopping= False, epsilon= 1e-08, hidden_layer_sizes= (10, 3), learning_rate= ‘constant’, learning_rate_init= 0.001, max_fun= 15000, max_iter= 200, momentum= 0.9, n_iter_no_change= 10, nesterovs_momentum= True, power_t= 0.5, random_state= 0, shuffle= True, solver= ‘lbfgs’, tol= 0.0001, validation_fraction= 0.1, verbose= False, warm_start= False

KNN, K-nearest neighbor; LR, logistic regression; MLP, multilayer perceptron; RF, random forest; XgBoost, eXtreme gradient boosting.

Table 4.

Comparison of the parameters of models in assessing the sleep quality of IBD patients.

Models	AUC	Accuracy	Sensitivity (recall)	Specificity	Precision	F1 score	Brier score
LR	0.75	0.73	0.58	0.88	0.83	0.68	0.19
RF	0.75	0.69	0.65	0.73	0.70	0.68	0.20
KNN	0.71	0.66	0.64	0.68	0.66	0.65	0.24
XgBoost	0.73	0.69	0.60	0.77	0.72	0.65	0.21
MLP	0.74	0.72	0.64	0.79	0.75	0.69	0.20
Voting	0.76	0.74	0.63	0.85	0.81	0.70	0.19

Accuracy = (TP + TN)/(TP + TN + FP + FN). Sensitivity (Recall) = TP/(TP + FN). Specificity = TN/(TN + FP). Precision = TP/(TP + FP). F1 score = 2 * Precision * Recall/(Precision + Recall). Brier score (Y, P) = 1/n *  ${\displaystyle \sum }_{i=1}^n{(P_i-Y_i)}^2$ .

The bold entries in the table are the maximum value in each column.

AUC, area under the curve of ROC; FN, false negative; FP, false positive; IBD, inflammatory bowel disease; KNN, K-nearest neighbor; LR, logistic regression; MLP, multilayer perceptron; n, number of predicted events; P, probability of model prediction; RF, random forest; TN, true negative; TP, true positive; XgBoost, eXtreme gradient Boosting; Y, actual probability of occurrence (no occurrence recorded as 0).

Figure 3.

The AUC of the ROC curve for each ML model in the evaluation of IBD-related sleep disorders.

AUC, area under the curve; IBD, inflammatory bowel disease; ML, machine learning.

Figure 4.

The calibration curves of the ML models, including MLP calibration curve (a), XgBoost calibration curve (b), voting calibration curve (c), KNN calibration curve (d), RF calibration curve (e), and LR calibration curve (f).

KNN, K-nearest neighbor; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron; RF, random forest.

Figure 5.

The DCA curves for the voting model.

DCA, decision curve analysis.

Feature importance and ML model interpretations

We calculated the feature importance of the optimal voting model (Figure 6). Depression is the most important feature in this model, followed (in order) by anxiety, age, outpatient diagnosis, course of disease, sex, severity, surgery, and first visit. To demonstrate the utility of the voting model in a real clinical situation, we used the LIME algorithm to interpret predictions for two representative cases: an IBD patient with sleep disorders and one without (Figure 7). The LIME analysis for the patient with sleep disorders (predicted probability: 0.92) identified the top contributing features to this prediction (Figure 7(a)). Depression (weight: +0.37), anxiety (+0.18), older age (+0.07), outpatient diagnosis (+0.07), male (+0.04), and an active or more severe stage of the disease (+0.01) were the positive drivers. While no surgical history (−0.06), not first visit (−0.03), and shorter disease duration (−0.03) reduced the risk. This aligns with clinical understanding: psychiatric comorbidities are established risk factors for sleep disruption in IBD. Conversely, for the patient without sleep disorders (predicted probability: 0.40), LIME highlighted surgical history (+0.07), outpatient diagnosis (+0.06), and longer disease duration (+0.05) as the top three positive contributors, but these were outweighed by protective factors: absence of depression (−0.37), absence of anxiety (−0.18), not first visit (−0.04), and female sex (−0.03; Figure 7(b)). In both cases, the voting model’s predictions and LIME’s interpretability aligned perfectly with the clinical realities. Therefore, local interpretation using the LIME algorithm has significant clinical value for aiding clinicians in assessing the risk of sleep disorders among IBD patients.

Figure 6.

The feature importance in order of the voting algorithm model.

Figure 7.

The voting model interpretation based on Lime algorithm in the evaluation of IBD patients without or with sleep disorders, positive sample (a) and negative sample (b).

IBD, inflammatory bowel disease.

Discussion

Principal results

In this multicenter study, we constructed several commonly used ML models for the assessment of IBD-related sleep disorders. To the best of our knowledge, this is the first study to undertake the evaluation of IBD-related sleep disorders based on an interpretable ML approach. After validating and comparing the models, we found that the voting model had the best performance and could be used for risk assessment in relation to whether patients with IBD develop associated sleep disorders. The model has greater utility and accuracy and can be used as a tool for clinical workers to perform risk assessments of IBD patients, enabling early identification and further intervention to improve the QOL of patients.

Comparison with prior work

The main symptoms of IBD patients are chronic recurrent abdominal pain and diarrhea combined with extraintestinal manifestations such as psychiatric disorders and sleep disorders. It has been shown that sleep disorders have a bidirectional relationship with IBD, ⁴⁴ which further leads to impaired QOL for patients. Therefore, targeting the improvement of sleep quality may serve as an effective intervention to improve the outcomes of patients with IBD. In recent years, several studies have focused on analyses of the correlation of sleep disorders with age, gender, anxiety, depression, clinical characteristics, disease severity, and QOL of IBD patients.^45–47 In the current study, we used ML for the first time to identify predictors of sleep disorders associated with IBD patients. After screening, a total of nine significant factors were identified, which were (in descending order of importance) depression, anxiety, age, outpatient diagnosis, course of disease, sex, severity, surgery, and first visit.

Depression and anxiety are the two most important risk factors. Depression and anxiety often coexist, and approximately 85% of patients with depression experience significant symptoms of anxiety, while as many as 90% of patients with anxiety disorders experience depression. ⁴⁸ The occurrence of depression or anxiety is more common in patients with IBD, with prevalence rates ranging from 29% to 35% during remission and up to 60%–80% during activity. ⁴⁹ Depressive or anxiety symptoms observed in patients with IBD may not be just related to disease behavior, but may be comorbidities that often exacerbate pre-existing diseases. The mechanism of action of the occurrence of anxiety and depression in IBD patients may include pro-inflammatory cytokines, the brain–gut axis, intestinal microbes, the nitric oxide pathway, genetic susceptibility, and many other modalities. Meanwhile, recurring intestinal symptoms, as well as prolonged and frequent outpatient and hospitalization treatments, bring great mental stress.^50–53 It has been shown that disease severity in IBD is an independent predictor of depressive symptoms, and patients presenting with symptoms such as pain have significantly higher depression scores than UC patients without associated symptoms. ⁵⁴ There is limited research on the order of occurrence and the degree of correlation between IBD and depression and anxiety, and the results of the available studies suggest that the relationship between IBD and depression and anxiety appears to be bidirectional, with an increased risk of depression and anxiety before and after the diagnosis of IBD. ⁵⁰ The incidence of depression and anxiety can lead to the recurrence or exacerbation of IBD through a variety of potential mechanisms, which can lead to the development or exacerbation of sleep disorders in patients. The increased frequency of nocturnal diarrhea, worsening of abdominal pain, and systemic glucocorticoid use when the patient is in the active phase of the disease can cause alterations in the patient’s sleep patterns.^55,56 According to available studies, nearly half of patients in clinical remission also have impaired sleep quality, which may be related to the presence of subclinical inflammation in patients. ¹⁴ At the same time, the stress and concern caused by the disease can put patients in a state of hyperarousal and vulnerability to external stimuli. ⁵⁷ In addition to this, there is a direct relationship between underlying or diagnosed mood disorders and sleep disorders. Studies have shown that there is a bidirectional relationship between sleep disorders and anxiety and depression. Most patients with depression or anxiety have clinical symptoms that include sleep disorders, including poor sleep quality, insufficient sleep quantity, and altered sleep pattern. Many anxiety disorders and depressive disorders have now included sleep disorders as one of the diagnostic criteria, ⁵⁸ and severe sleep disorders can also contribute to anxiety and depression.

In the present study, age was also found to be one of the more significant predictors of comorbid sleep disorders in patients with IBD, as shown by the fact that the older the age of the patient with IBD, the higher the likelihood of predicting the development of sleep disorders. With increasing age, sleep patterns will change, including changes in sleep pattern, decreasing sleep duration, and changes in the ratio of slow-wave sleep to rapid eye movement sleep, causing a decrease in their QOL. ⁵⁹ Due to the inherent special characteristics of elderly IBD patients, there is also a lack of efficacy trials on elderly patients. At the same time, elderly patients with IBD tend to have multiple comorbidities compared to younger IBD patients, and older adults have worse immune functions. Age-related physiologic changes may alter the pharmacokinetics and metabolism of drugs, and the therapeutic process tends to be more prolonged and complex, ⁶⁰ which may have an impact on the patient’s sleep quality. Based on the results of permutation importance, we found that outpatient diagnosis and a longer course of illness also tend to favor the development of sleep disorders. There are fewer studies on the direct relationship between outpatient diagnosis and co-morbid sleep disorders in IBD patients. Outpatient diagnosis patients may be in the active phase of the disease, and symptoms such as abdominal pain, diarrhea, and blood in the stool may interfere with sleep quality. In addition to this, outpatient patients may still lack a comprehensive and correct understanding of IBD, and may be suffering from a variety of psychological problems, such as anxiety and fear, which may also have a negative impact on their sleep quality. With the prolongation of IBD, long-term and frequent abdominal pain, diarrhea, and other discomforts may cause patients to suffer both physical and mental torture, which may lead to psychological problems such as anxiety, depression, and insomnia. ⁶¹

Both UC and CD, as two subtypes of IBD, exhibit a high prevalence of sleep quality disturbances (48.43% UC vs 49.41% CD), while the underlying mechanisms appear to differ. CD patients demonstrated stronger associations between penetrating disease and sleep fragmentation, likely mediated by chronic abdominal pain. ¹⁴ In contrast, UC patients more frequently reported nocturnal bowel movements, which significantly correlated with reduced sleep efficiency. ⁶² These findings suggest that while both UC and CD patients require comprehensive sleep evaluation, optimal management strategies may differ. CD patients may benefit most from aggressive pain control and earlier biologics escalation in those with penetrating disease, whereas UC patients might achieve greater sleep improvement through targeted management of nocturnal bowel frequency. Future studies incorporating objective sleep measures and detailed phenotyping are needed to further elucidate these subtype-specific relationships.

Many researchers are aware of sleep problems in patients with IBD, and several studies have reported interactions between patients’ sleep disorders and disease activity, as well as their correlations with QOL, parenteral manifestations, and medications, but the findings are not entirely consistent.^15,45 Traditional disease risk prediction is mainly based on Cox proportional risk regression models and LR models, which are widely used, but there is still room for improvement in terms of both prediction accuracy and model interpretability. In recent years, feature selection and supervised learning models in the field of AI machine learning have been increasingly applied to the disease prediction problem, which can improve the interpretability of prediction models, and some emerging methods show better prediction performance. No studies have been reported on the use of predictive models for IBD-related sleep disorders. In this study, the AUC of the voting algorithm was found to reach 0.76, which indicates that the voting model has a high predictive performance in the context of predicting sleep disorders. Overall, the application of the voting model can be helpful in identifying and assessing IBD-related sleep disorders, with good predictive performance.

Clinical translation of predictive modeling

While previous studies have established isolated associations between sleep disorders and factors like depression or disease activity, our ML model transforms these known relationships into a clinically actionable tool through three key innovations. First, it replaces population-level odds ratios with personalized risk quantification, enabling precision resource allocation. Second, the model’s Lime-based explanations reveal contextual interactions between factors and capture dynamic risk patterns unrecognized by traditional analyses. Third, the voting model’s implementation flexibility allows adaptation to real-world settings through electronic health records (EHR) integration, local data retraining, and threshold adjustment via decision curve analysis. This represents a paradigm shift from merely documenting risk factors to operationalizing them through automated high-risk patient identification, case-specific decision support, and continuous learning from clinical workflows—effectively bridging the gap between epidemiological knowledge and bedside practice in IBD sleep management.

While the predictive performance of our model (AUC 0.76) demonstrates technical validity, its true clinical value lies in actionable implementation pathways designed through clinician feedback. The model enables: (1) Automated risk stratification—integrating with hospital EHRs to flag high-risk patients for immediate sleep assessment and screening during routine IBD visits; (2) Precision intervention—guiding treatment choices by revealing modifiable risk drivers (e.g., prioritizing anxiety management when LIME shows >15%); and (3) Economic efficiency—providing some relief and reduce recurrence, saving avoidable hospitalization costs. Implementation requires minimal infrastructure (pre-built API in Code) and aligns with existing IBD care protocols, transforming predictive analytics into measurable clinical impact.

We have developed a roadmap for prospective validation and real-world implementation of our predictive model, consisting of three key phases. First, we will conduct a 12-month multicenter prospective validation study that will include enrolling a number of IBD patients from Chinese hospitals to evaluate real-time performance (AUC stability, sensitivity/specificity drift) and clinician adherence to recommendations. Second, we will launch a real-time clinical pilot integrating the model as a Fast Healthcare Interoperability Resources-based module into three hospital EHR systems, featuring automated alerts for high-risk cases (PSQI >6) and measuring time-to-intervention metrics. Third, a planned hybrid effectiveness-implementation trial will use a randomized controlled trial design to compare model-guided versus usual care, with primary endpoints of IBD flare reduction (target hazard ratio (HR) ⩽0.7) and cost-effectiveness analysis. Implementation strategies include clinician-facing EHR alerts with one-click referrals and patient-facing SMS reminders with risk-stratified education. Success metrics focus on clinical utility while maintaining safety. This comprehensive translation plan bridges our technical development to measurable clinical impact through rigorous prospective evaluation and adaptive implementation.

While this study derived its model from a Chinese IBD cohort, it has potential availability in other populations as well. First, key predictors (e.g., depression, disease activity) demonstrate consistent associations with sleep disturbances in multinational IBD studies.^63,64 Second, the PSQI, PHQ-9, and GAD-7 used here have been validated across diverse populations. Still, the model has some limitations due to potential genetic differences and health care disparities. This relies on the inclusion of a wider range of raw data in the future in order to adapt the model to populations with different characteristics.

Limitations

This study featured some limitations. First, the features included in this study were not comprehensive, such as the type of disease, the frequency of strenuous exercise, the number of night shifts, and inflammatory markers, which have been shown in some studies to be strongly associated with sleep disorders in patients with IBD.^9,63 In addition, in this study, the PSQI scale was used as the outcome, which is a single referenced indicator and lacks an objective measure of sleep quality for patients. ⁶⁴ Secondly, the data and information of this study came from 42 hospitals in 22 provinces in China, and 2478 IBD patients were ultimately enrolled. The sample size was relatively small and came from a single source, and factors such as other races were not considered. There were limitations in the applicability of the prediction model. Another limitation of this study is that we did not perform separate analyses for UC and CD and can’t develop more tailored management strategies for each IBD subtype. The questionnaire-based methodology employed in this study may be subject to reporting biases, including potential under- or over-reporting of symptoms due to social desirability bias, recall bias, and interpretation variability. Therefore, there is still room for improvement in the construction of a prediction model for IBD co-morbid sleep disorders, and this study may provide a reference for subsequent researchers.

Conclusion

In conclusion, we developed and validated a risk assessment model for IBD-related sleep disorders for the first time using an interpretable ML approach. The results show that the voting algorithm presented the best predictive performance among multiple models. A more comprehensive consideration of inclusion features and outcome indicators is still needed to improve accuracy in the future, and the sample size and source can be expanded to support the broad applicability of assessment models.

Appendix

Abbreviations

AI artificial intelligence

CD Crohn’s disease

GAD-7 Generalized Anxiety Disorder 7-item Scale

IBD inflammatory bowel disease

PHQ-9 Patient Health Questionnaire-9

PSQI Pittsburgh Sleep Quality Index

QoL quality of life

UC ulcerative colitis