Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study

Abstract

Background

Post-stroke fatigue, as one of the long-lasting physical and mental symptoms accompanying stroke survivors, will seriously affect the daily living ability and quality of life of stroke patients.

Objective

The aim of this study was to develop machine learning (ML) algorithms to predict early post-stroke fatigue among patients with stroke.

Methods

A longitudinal study of 702 patients with stroke followed for 3 months. Twenty-three clinical features were obtained from medical records and questionnaires before discharge. Early post-stroke fatigue was assessed using the Fatigue Severity Scale. The dataset was randomly divided into a training group (70%) and an internal validation group (30%), applied oversampling, 10-fold cross-validation, and grid search to optimize the hyperparameter. Feature selection using the Least Absolute Shrinkage and Selection Operator (LASSO) regression. Sixteen ML algorithms were performed to predict early post-stroke fatigue in this study. Accuracy, precision, recall, F1 score, area under the receiver operating characteristic curve (AUC), and brier score were used to evaluate the models performance.

Results

Among the 16 ML algorithms, the Bagging model was the optimal model for predicting early post-stroke fatigue in patients with stroke (AUC = 0.8479, accuracy = 0.7518, precision = 0.5741, recall = 0.7209, F1 score = 0.6392, brier score = 0.1490). The feature selection based on LASSO revealed that risk factors for early post-stroke fatigue in patients with stroke included anxiety, sleep, social support, family care, pain, depression, neural-functional defect, quit/no drinking, balance function, type of stroke, sex, heart disease, smoking, and hemiplegia.

Conclusions

In this study, the Bagging model proved to be effective in predicting early post-stroke fatigue.

Keywords

machine learning stroke post-stroke fatigue prediction risk factors

Introduction

Stroke is the second leading cause of death in the world¹ and the first cause of disability and death in China.² Stroke, also known as “cerebrovascular accident,” is an acute cerebrovascular disease due to the sudden rupture of blood vessels in the brain or a type of disease that causes brain tissue damage due to blood obstruction that prevents blood from flowing into the brain, mainly including ischemic stroke and hemorrhagic stroke.³ Stroke is often accompanied by physical dysfunction, speech dysfunction, consciousness disturbance, and emotional complications such as disorders,⁴ among which post-stroke fatigue is one of the common complications in stroke patients.

Post-stroke fatigue is a kind of persistent pathological fatigue that cannot be alleviated by rest.⁵ Post-stroke fatigue generally starts in the acute phase of stroke patients and lasts for 2 years or more.⁶ Post-stroke fatigue is divided into early post-stroke fatigue(within 3 months) and chronic post-stroke fatigue(more than 3 months).^7,8 The 2 types of fatigue may be distinct but are not considered mutually exclusive. Previous studies showed that early post-stroke fatigue might be a predictor of chronic post-stroke fatigue.^8,9 Therefore, predicting early post-stroke fatigue not only prevents or reduces the occurrence of post-stroke fatigue but also effectively prevents the occurrence of chronic post-stroke fatigue. Pedersen et al¹⁰ found in a prospective study of 430 Swedish patients with ischemic stroke that 80% of the patients still had post-stroke fatigue 7 years after the onset of stroke. A study by Lin and Chen¹¹ showed that the incidence of post-stroke fatigue in China ranged from 25% to 85%. Additionally, post-stroke fatigue can easily lead to a decline in physical tolerance,¹² which affects the rehabilitation training process of patients, reduces the rehabilitation effect of stroke patients, and increases the care burden of patients’ families.¹³ Consequently, early prediction of post-stroke fatigue in patients with stroke is crucial.

Standard assessment of these fatigue symptoms typically utilizes some scales¹⁴ such as the Fatigue Severity Scale (FSS), Fatigue Impact Scale, Fatigue Assessment Scale, Mental Fatigue Scale, Neurological Fatigue Index for Stroke, and so on. Although these scales are widely used in clinical practice, the focus of their assessment may be different, and it is difficult to guarantee the accuracy of the assessment of post-stroke fatigue. Kou et al¹⁵ pointed out in the evidence summary study that using the FSS to evaluate post-stroke fatigue was recommended. Thus, the FSS was used to evaluate early post-stroke fatigue in this study.

Machine learning (ML) algorithms have increasingly been employed in the medical field to develop predictive models for complex diseases, including post-stroke fatigue. Luzum et al¹⁶ predicted chronic post-stroke fatigue in patients with stroke and found that Random Forest model was the optimal model (area under the receiver operating characteristic curve [AUC] = 0.79). Their study provided an effective tool for predicting chronic post-stroke fatigue. Recently, numerous scholars have started examining the predictors of post-stroke fatigue. Various studies have endeavored to develop ML algorithms to predict post-stroke fatigue based on biochemical indexes. Alterations in the frontal thalamic striatal system or inflammatory processes may cause post-stroke fatigue.¹⁷ However, the primary causes of post-stroke fatigue in patients remain largely unknown. Researchers have discussed several factors that may be associated with post-stroke fatigue, including sex, depression, sleep, neurological, or biological factors, among others.^18
-20 Wu et al²¹ predicted post-stroke fatigue at 6 and 12 months by collecting C-reactive protein (CRP) from stroke patients at the Royal Hospital of Edinburgh, United Kingdom, and found that CRP level had no significant predictive effect on fatigue at 6 or 12 months. However, Liu et al²² found that CRP in patients with stroke in China could predict the post-stroke fatigue of patients for 6 months. In addition, Ren et al²³ found that serum uric acid level at admission could also predict post-stroke fatigue. At the same time, some studies have focused on the predictive effect of psychological factors on post-stroke fatigue in patients with stroke. Kliem et al²⁴ showed that self-reported cognitive and psychiatric symptoms at 3 months could predict daytime sleep and fatigue at 12 months. A recent study found that depression was one of the predictors of post-stroke fatigue.¹⁶ Additionally, Cui et al²⁵ found that social support could also affect post-stroke fatigue. Nevertheless, few studies have fully considered physiological, psychological, and social factors to predict early post-stroke fatigue.

Therefore, to address this gap, the purpose of the current study was to explore physiological, psychological, and social predictors of early post-stroke fatigue in patients with stroke using 16 ML algorithms. We further analyzed important risk factors of early post-stroke fatigue to facilitate the clinical application of the developed optimal model.

Methods

Study Design

The study was conducted by an interdisciplinary research team consisting of scholars in nursing, rehabilitation, and statistics. This longitudinal study was conducted from June 2023 to May 2024. The study followed the Transparent Reporting of a Multivariate Prediction Model for Individual Prognosis or Diagnosis reporting guidelines.²⁶ The study was carried out in accordance with the Declaration of Helsinki and approved by the Medical Ethics Committee of the China Rehabilitation Research Center (No: 2023-114-01). All participants were informed and consented to participate in the study.

Participants

Participants were recruited from a tertiary A hospital in Beijing, China. This study included patients who met the diagnostic criteria for stroke (diagnosed by computed tomography or magnetic resonance imaging of the head) established at the Fourth National Conference on Cerebrovascular Diseases in 1995.²⁷ Patients were excluded if they had a stroke for more than a month; were not first-time stroke; had cancer, systemic lupus erythematosus, Parkinson’s disease, or other known fatigue diseases; scored below 24 on the Mini-Mental State Examination Scale²⁸; had speech or communication disorders and were unable to communicate; were diplegia; did not complete the 3-month follow-up; and scored below 45 on the activities of daily living.²⁹ According to the sample size empirical formula,³⁰ $N = \frac{\begin{array}{l} 10 \times (1 + Invalid response rate) \\ \times Events per predictor variable \end{array}}{Prevalence}$ . The previous meta-analysis results showed that the prevalence of post-stroke fatigue in Chinese stroke patients was about 46%, and 5 to 15 predictor variables were estimated in the training model.³¹ To avoid extreme outliers of responders or non-responders, the invalid response rate was set to 20%. Thus, the sample size of the training set was 131 to 392. The internal validation set was 57 to 168 patients according to a 7:3 ratio.

Measures

Predictor Variables

In order to make the predictors of early post-stroke fatigue more comprehensive, a Biopsychosocial Model was used to collect predictive variables. This model, created by George Engel in 1977,³² included biological (eg, genetic factors, physiological, and biochemical factors), psychological (eg, anxiety and depression), and social (eg, social activities and interpersonal relationships) factors. In this study, factors such as motor function and balance function were taken as biological factors, anxiety and depression as psychological factors, and social support and family care as social factors. Twenty-three predictors were collected at baseline, including sex, age, body mass index (BMI), smoking, drinking, heart disease, hypertension, diabetes, type of stroke, hemiplegia, motor function, balance function, neural-functional defect, sleep, pain, serum uric acid, CRP, homocysteine, Cerebral regional oxygen saturation (rScO₂), anxiety, depression, social support, and family care. Serum uric acid, CRP, and homocysteine were the last values recorded before discharge.

Motor Function

Fugl-Meyer et al³³ proposed the Fugl-meyer Assessment Scale, primarily based on Brunnstrom’s views, to evaluate motor function of stroke patients in their upper and lower limbs. The upper limb assessment included 10 dimensions and 33 items, and the lower limb assessment included 7 dimensions and 17 items. The 3-level scoring method was adopted, which were recorded as 0, 1, and 2 points respectively. The total score on the scale was 100 points. A total score of less than 50 was classified as grade I, indicating serious dysfunction of the affected limb; 50 to 84 was divided into grade II, indicating that there were obvious movement disorders of the affected limb; 85 to 95 was classified as grade III, indicating moderate movement impairment of the affected limb; 96 to 99 was classified as grade IV, indicating mild motor impairment of the affected limb.

Balance Function

The Fugl-meyer Balance Scale has evaluated the balance function of stroke patients with hemiplegia, including 7 items³⁴: unsupported sitting, “wingspreading” response on the healthy side, “wingspreading” response on the affected side, standing with support, standing without support, standing on the healthy side, and standing on the affected side. Each item was scored in 3 levels of 0 to 2 points, with the highest score being 14 and the lowest score being 0. The lower the score, the more severe the balance dysfunction.

Neural-Functional Defect

The National Institute of Health Stroke Scale (NIHSS) could be used to assess the neural-functional defect in stroke patients.³⁵ The Chinese version of NIHSS was translated by Hou et al³⁶ and demonstrated good reliability. Cronbach’s α coefficient was .796. The score ranged from 0 to 42 points. The higher the score, the more serious the neural-functional defect. 1 to 4 was classified as mild stroke, 5 to 15 as moderate stroke, 16 to 20 as moderate to severe stroke, and 21 to 42 as severe stroke.

Sleep

The Pittsburgh Sleep Quality Index Scale was proposed in 1989 by Buysse et al³⁷ and carried out in Chinese by Liu et al,³⁸ mainly assessed sleep conditions. The Cronbach’s α coefficient was .842, indicating acceptable internal consistency. The scale consisted of 19 self-rated items and 5 other rated items. It mainly included 7 dimensions (sleep latency, sleep efficiency, sleep quality, sleeping time, hypnotic drugs, sleep disorders, and daytime function), and each dimension was scored according to 0 to 3 points. The total score ranged from 0 to 21 points. The higher the score, the worse the sleep quality.

Pain

The Number Rating Scale was adopted to evaluate the pain level of patients.³⁹ Pain level was measured on a scale of 0 to 10,⁴⁰ with higher scores indicating more severe pain.

Anxiety

The Hamilton Anxiety Scale was developed by Hamilton in 1959 to assess anxiety, and it included 14 items.⁴¹ All items were scored on a 5-point scale ranging from 0 to 4. The asymptomatic score was 0, 1 was mild, 2 was moderate, 3 was severe, and 4 was extremely severe. The total score of less than 7 indicated no anxiety. The Cronbach’s α coefficient was .93.⁴²

Depression

The Hamilton Depression Scale was developed by Hamilton⁴³ in 1960 to assess depression, and it included 17 items. Items 1 to 9 were graded on a 5-point scale of 0 to 4. Criteria at all levels: 0 was asymptomatic, 1 was mild, 2 was moderate, 3 was severe, and 4 was extremely severe. Items 10 to 17 used a 3-level evaluation method with 0 to 2 points, and the grading criteria were 0 as asymptomatic, 1 as mild to moderate, and 2 as severe. A total score of less than 7 was considered normal. The Cronbach’s α coefficient was .99.⁴²

Social Support

In 1986, Xiao⁴⁴ designed the Social Support Rating Scale, which mainly assessed the social support level of patients. The scale included 3 dimensions of objective support, subjective support, and social support utilization, with a total of 10 items and a total score of 12 to 66 points. A higher score indicated a higher level of social support. The Cronbach’s α coefficient of the scale was .783, indicating good internal consistency.⁴⁵

Family Care

The Family APGAR Index was compiled by Smilkstein et al,⁴⁶ in 1978 to assess family function status. The Cronbach’s α coefficient of the scale was.813, indicating good reliability. The Chinese version was introduced by Lv and Gu,⁴⁷ in 1995 to assess the satisfaction of family function. The scale included 5 items: adaptation, partnership, growth, affection, and resolve. Three possible answers were allowed (“Almost always,” “Sometimes,” and “hardly ever”), and the item score ranged from 0 to 2 points. A total score of 0 to 10 points and higher scores indicated better family care.

Outcome Definition

According to the best evidence summarized by Kou et al,¹⁵ it was recommended to use FSS for fatigue assessment of patients. FSS was developed by Krupp et al⁴⁸ and translated into Chinese by Wu and Wang⁴⁹ to evaluate post-stroke fatigue. The scale included 9 items, which adopted a Likert 7-point scale. The average score of the 9 items was the score of fatigue. Cronbach’s α coefficient was .929, indicating good reliability. In this study, the FSS was used to diagnose early post-stroke fatigue in patients with stroke after 3 months of follow-up. When the score was greater than or equal to 4, it was considered early post-stroke fatigue. In contrast, early post-stroke fatigue was not considered.

Data Collection

A total of 702 patients were examined in a tertiary A hospital in Beijing, China, between June 2023 and January 2024. Initial data were collected through the use of medical records and face-to-face questionnaires. A total of 561 patients successfully completed a 3-month telephone follow-up, with a completion rate of 79.91%. Among these patients, 392 were in the training group, and 169 were in the internal validation group. A total of 202 patients were subjected to external validation over the period from February 2024 to May 2024. The valid sample size of the external validation set was 141 patients (69.80%).

Statistical Analysis

IBM Statistical Package for Social Sciences v.21, Python v.3.12, and Visual Studio Code v.1.90 were used for data analysis. Variables such as age and BMI were represented by mean ± standard deviation, while count and percentage represented variables such as sex and type of stroke. If the P value was <.05, statistical difference was considered.

Data Pre-Processing

In this study, the missing values were simply filled in by means of mean or median. The predictive variables of drinking and smoking were set as dummy variables. Because of the small sample size, SMOTE oversampling was used to deal with the unbalanced data.⁵⁰ The min-max normalization technique was used to scale the data to a specific range, which was beneficial in improving the performance and accuracy of the model.⁵¹

Model Construction

In this study, Decision Tree, Random Forest, Logistic Regression, Stochastic Gradient Descent (SGD), K-Nearest Neighbor, Gaussian Naive Bayes (Gaussian NB), Bernoulli naive Bayes (Bernoulli NB), Multinomial naive Bayes (Multinomial NB), Support Vector Classification, Quadratic Discriminant Analysis, Extremely Randomized Trees, Linear Discriminant Analysis, Adaptive Boosting (AdaBoost), Bagging, Gradient Boosting, and eXtreme Gradient Boosting (XGBoost) were used to establish models of early post-stroke fatigue assessed at a 3-month follow-up.⁵² T-test, Mann–Whitney U test, and Chi-square test were used to analyze the differences between the model and external validation groups. Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for feature selection to screen out the features that contributed the most to the prediction model. After feature selection, a set of optimal hyperparameters was selected through 10-fold cross-validation of the training group.

Model Performance

The validation group was used to evaluate the models’ performance in predicting early post-stroke fatigue in patients. In this study, the models were evaluated in terms of distinction (accuracy, recall, precision, F1 score, and AUC) and calibration (Brier score). Accuracy refers to the proportion of the number of samples correctly predicted to the total number of samples. The recall refers to the proportion of the actual positive samples predicted to be positive samples in the actual total samples. The precision refers to the proportion of the predicted positive samples in the actual positive samples to the total positive samples in the prediction model. The higher the values of the above 3, the better the performance of the model. The F1 score is an indicator to measure precision and recall. The value of F1 ranges from 0 to 1. The closer the F1 score is to 1, the better the comprehensive performance of the model. The AUC ranges from 0.5 to 1, and the higher the score, the better the predictive model performance. The brier score ranges from 0 to 1 and assesses the difference between the models’ categorical predictions and the actual results, with a lower score indicating a higher likelihood of calibration. When various indexes are inconsistent, AUC is the main index of model performance.

Results

Population Statistics

From June 2023 to January 2024, 900 stroke patients were investigated in this study, among which 198 patients did not meet the inclusion and exclusion criteria and were therefore excluded (Supplemental Figure 1). A total of 702 patients were included in this study, 561 of whom completed a 3-month follow-up and were divided into the training group and the internal validation group according to a ratio of 7:3. From February to May 2024, a total of 141 stroke patients participated in the study as an external validation group and completed follow-up. Table 1 summarizes the physiological, psychological, and sociological predictors and compares the baseline data of the modeling group with that of the validation group, finding no statistical difference between the 2 groups (P > 0.05). In this study, the incidence of early post-stroke fatigue was 30.3%. The score for patients with early post-stroke fatigue was 5.22 ± 0.81, while those without early post-stroke fatigue was 2.15 ± 0.92.

Table 1.

Demographic Characteristics of Patients With Stroke in the Training Group, Internal Validation Group, and External Validation Group.

Variables	Modeling group (N = 561)		External validation group (N = 141)	Statistic	P
Variables	Training group (N = 392)	Internal validation group (N = 169)	External validation group (N = 141)	Statistic	P
Age	58.64 ± 13.80	59.00 ± 12.95	59.60 ± 11.85	−0.332^b	.740
Sex (%)				0.130^c	.718
Female	135 (34.4)	49 (29.0)	44 (31.2)
Male	257 (65.6)	120 (71.0)	97 (68.8)
BMI, Kg/m²	23.37 ± 1.75	23.47 ± 1.60	23.26 ± 1.95	0.854^a	.393
Type of stroke (%)				0.038^c	.846
Ischemic stroke	253 (64.5)	110 (65.1)	90 (63.8)
Hemorrhagic stroke	139 (35.5)	59 (34.9)	51 (36.2)
Hemiplegia (%)				0.187^c	.665
Left hemiplegia	189 (48.2)	89 (52.7)	67 (47.5)
Right hemiplegia	203 (51.8)	80 (47.3)	74 (52.5)
Smoking (%)				2.201^c	.333
No	194 (49.5)	84 (49.7)	63 (44.7)
Quit smoking	55 (14.0)	19 (11.2)	25 (17.7)
Yes	143 (36.5)	66 (39.1)	53 (37.6)
Drinking (%)				0.573^c	.751
No	189 (48.2)	77 (45.6)	66 (46.8)
Quit drinking	23 (5.9)	11 (6.5)	11 (7.8)
Yes	180 (45.9)	81 (47.9)	64 (45.4)
Heart disease (%)				0.003^c	.957
No	290 (74.0)	129 (76.3)	105 (74.5)
Yes	102 (26.0)	40 (23.7)	36 (25.5)
Hypertension (%)				0.155^c	.694
No	106 (27.0)	40 (23.7)	39 (27.7)
Yes	286 (73.0)	129 (76.3)	102 (72.3)
Diabetes (%)				0.607^c	.436
No	242 (61.7)	96 (56.8)	90 (63.8)
Yes	150 (38.3)	73 (43.2)	51 (36.2)
rScO₂, %	66.20 ± 3.47)	66.26 ± 3.23	66.45 ± 3.30	−0.715^a	.475
Serum uric acid, µmol/L	301.26 ± 89.52	299.53 ± 88.87	290.94 ± 86.90	1.172^a	.242
C-reactive protein, mg/L	4.26 ± 4.72	4.59 ± 5.06	5.04 ± 5.18	−0.970^b	.332
Homocysteine, µmol/L	12.05 ± 4.30	12.23 ± 4.78	11.35 ± 3.54	1.860^a	.063
Sleep	2.56 ± 2.54	2.63 ± 2.75	2.61 ± 2.69	−0.124^a	.901
Pain	0.37 ± 0.69	0.47 ± 0.72	0.40 ± 0.75	−0.074^a	.941
Motor function	63.55 ± 23.34	59.62 ± 22.61	58.16 ± 23.72	1.921^a	.055
Balance function	8.02 ± 2.96	7.64 ± 2.96	7.35 ± 3.07	1.958^a	.051
Neural-functional defect	4.76 ± 2.65	5.14 ± 2.54	5.13 ± 2.75	−1.026^a	.305
Anxiety	6.50 ± 5.82	7.07 ± 6.46	6.74 ± 6.29	−0.130^a	.897
Depression	7.84 ± 5.04	7.74 ± 5.12	7.79 ± 5.16	0.053^a	.957
Social support	39.96 ± 5.83	39.98 ± 6.42	39.33 ± 6.06	1.119^a	.264
Family care	7.03 ± 2.12	7.05 ± 2.25	6.68 ± 2.33	1.701^a	.089

T-test.

Mann–Whitney U test.

Chi-square test.

Feature Selection

In order to simplify the model and improve its generalization ability, we used LASSO with the introduction of L1 regularization algorithm (alpha [α]) to select important features. The reason why we used LASSO algorithm was that it was easy to screen out important variables that had a greater impact on dependent variables when there were a large number of predictors. As the α value increased, the features were compressed smaller, and when the features were compressed to 0, it meant that the variables were eliminated from the model. The later a variable was compressed to zero, the more influential it represented the variable. All the selected variables mentioned above were included in LASSO regression. On the basis of the optimal α parameter, the parameter α of LASSO regression is selected by using the 10-fold cross-validation method (Supplemental Figure 2). Using the minimum mean squared error, a dashed vertical line (α = .00364386) is drawn at the ideal value, and the important features includes anxiety, sleep, social support, family care, pain, depression, neural-functional defect, quit /no drinking, balance function, type of stroke, sex, heart disease, smoking, and hemiplegia (Supplemental Figure 3).

Model Construction

Supplemental Table 1 shows the results of 10-fold cross-validation of 16 models in the training group. The models superior to Logistic Regression in this study include Random Forest, SGD, GaussianNB, Extremely Randomized Trees, Linear Discriminant Analysis, AdaBoost, Bagging, Gradient Boosting, and XGBoost models, in which Bagging is the optimal model.

Model Performance

Figure 1 provides the performance of the models in internal validation. The results show that AUC ranges from 0.7389 to 0.8538, and the AUC value of Bagging model is the highest (AUC = 0.8538). External validation results in Figure 2 show that the Bagging model (AUC = 0.8479, accuracy = 0.7518, precision = 0.5741, recall = 0.7209, F1 score = 0.6392, and brier score = 0.1490) is also superior to other models. There was little difference between the predicted and actual results (brier score = 0.1490). Figures 3 and 4 show the ROC curves of the prediction model for internal and external validation, with Bagging performing the best overall. Therefore, considering the best degree of differentiation and calibration, Bagging is the best prediction model.

Figure 1.

Performance of the models in internal validation.

Figure 2.

Performance of the models in external validation.

Figure 3.

ROC of prediction model for internal validation.

Figure 4.

ROC of prediction model for external validation.

Model Application

We used Bagging algorithm to design an online web page. When the medical staff or patients input the values of the required variables online, they will get the predicted results (high-risk group for early post-stroke fatigue or low-risk group for early post-stroke fatigue) by clicking the prediction button (Figure 5). Once the results show that the patients are at high risk for early post-stroke fatigue, effective intervention measures should be taken early to control the development of post-stroke fatigue.

Figure 5.

Online page.

Discussion

In this study, the incidence of early post-stroke fatigue was 30.3%. This result is in line with the results of Lin and Chen,¹¹ indicating that the results of this study are reliable. This study constructed predictive models of early post-stroke fatigue by considering psychological, physiological, and social factors and developed an online web page of Bagging algorithm, which could provide a good predictive tool for medical staff or patients.

In this study, Bagging algorithm showed better performance compared with other models. Bagging algorithm is an ensemble meta-classifier, originally proposed by Leo Breiman in 1996, which consists of a set of basic classifiers applied to a random subset of the original data set.⁵³ The results of these classifiers are collected and a final prediction is drawn from them by voting. Mato-Abad et al⁵⁴ found that ML algorithm based on Bagging could better predict and diagnose the isolated syndrome. In addition, Xie et al⁵⁵ found that a precise identification model of patients with osteoporosis kidney-yang deficiency syndrome was conducted based on the rule ensemble method of Bagging combined with LASSO regression, and the selected key rules could better explain the identification process of kidney-yang deficiency syndrome and assist the patients with osteoporosis clinical differentiation of kidney-yang deficiency syndrome of traditional Chinese medicine. The above showed that Bagging algorithm had good performance in diagnosing and predicting diseases. Therefore, in the results of this study, Bagging algorithm performed best among the 16 ML models, which was reasonable.

In previous studies, many scholars have explored the prediction of post-stroke fatigue. Schnitzer et al⁵⁶ conducted a longitudinal study to predict post-stroke fatigue among stroke survivors in Gothenburg, but their study did not involve psychological and social predictors. Snaphaan et al⁵⁷ assessed psychological risk factors for early post-stroke fatigue in Dutch stroke patients, but the authors did not take into account social factors. Luzum et al¹⁶ took into account the prediction of post-stroke fatigue in patients with cognitive impairment. Their study provided a theoretical basis for future research into the prediction of post-stroke fatigue in patients with cognitive impairment, but this could complicate the prediction of post-stroke fatigue. In addition, their study used objective predictors (such as a handhold dynamometer measured grip-strength) to predict chronic post-stroke fatigue, which was a valuable approach to consider.

In order to identify early post-stroke fatigue, our study creatively considered physiological, psychological, and social factors to predict early post-stroke fatigue in patients with stroke for 3 months. XGBoost in this study had good accuracy and precision(accuracy = 0.8085, precision = 0.7105), which meant that it could accurately predict the occurrence of early post-stroke fatigue. In addition, the recall of the Decision Tree was good(recall = 0.8837), suggesting that the model predicted more true early post-stroke fatigue, which was important for early identification of early post-stroke fatigue. Compared with Logistic Regression, the F1 value of Bagging model was higher(F1 = 0.1490), but it achieved the highest AUC value, indicating that Bagging model has good robustness and performance. Finally, we selected Bagging algorithm as the optimal model (AUC = 0.8479), providing a comprehensive and effective prediction tool for early detection of high-risk groups of early post-stroke fatigue in the future.

The results of this study showed that anxiety, sleep, social support, family care, pain, depression, neural-functional defect, quit/no drinking, balance function, type of stroke, sex, heart disease, smoking, and hemiplegia were important features of Bagging algorithm in predicting early post-stroke fatigue. Similar to the results of this study, Wu et al²¹ found that CRP could not predict post-stroke fatigue. Mood disturbances are common in stroke patients. Early post-stroke fatigue was related to anxiety and depression.¹⁰ Future studies need to further study whether they interact with each other and how they affect early post-stroke fatigue. White et al⁵⁸ found that sleep was related to post-stroke fatigue, which might be related to the fact that sleep was an important link in human recovery and repair. Regular sleep was an important part of post-stroke fatigue management. Improving sleep quality could alleviate post-stroke fatigue in patients with hemiplegia. Stroke patients generally had limb dysfunction and difficulties in social readjustment, and good social and family support could not only help patients with social readjustment but also reduce patients’ psychological pressure and relieve post-stroke fatigue. Similarly, post-stroke fatigue was highly correlated to pain in the chronic stage of stroke (r = .39).⁵⁹ This is consistent with the results of this study on pain as a predictor. Guan et al⁶⁰ found that physical function and stroke characteristics are related to post-stroke fatigue, which was similar to the results of this study. In addition, their study also found that demographic data such as sex, smoking, and heart disease were also associated with post-stroke fatigue. To sum up, the predictive factors selected in this study were accurate and reliable. In the future, further studies will be conducted to explore the mechanism of predictive factors on post-stroke fatigue to provide the theoretical basis for early prevention of post-stroke fatigue.

Limitations

The study has some limitations. First, we only predicted early post-stroke fatigue 3 months after stroke onset and did not further follow up patients with post-stroke fatigue. In addition, we had to acknowledge the limitations of not measuring acute post-stroke fatigue at baseline. We ignored the predictive effect of acute post-stroke fatigue on early post-stroke fatigue. The relationship between acute post-stroke fatigue, early post-stroke fatigue, and chronic post-stroke fatigue could be explored in the future. Second, Chinese localized stroke diagnostic criteria were used instead of international diagnostic criteria in this study, which led to limitations in the internationalization and universality of the findings. In addition, we acknowledged that large-scale and multi-center external validation was better. Therefore, it is necessary to carry out large-scale studies using international tools in different regions and centers, which may help improve our model’s performance indicators further. Finally, this study did not explore the potential interaction between predictors, but it was a novel perspective to make the results more rigorous and definitive. In future studies, it is necessary to consider the potential interaction between the predictors.

Conclusions

In this study, we used data collected from medical records and questionnaires to develop an effective model for predicting early post-stroke fatigue using a ML algorithm. Our model reached high levels of AUC 3 months after predicting the onset of early post-stroke fatigue, demonstrating its ability to detect early in people at high risk for early post-stroke fatigue. Our online page helps identify patients at risk early and accurately, promotes the strategic goals of intervention, and improves patients’ early post-stroke fatigue.

Supplemental Material

sj-docx-4-nnr-10.1177_15459683251329893 – Supplemental material for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study

Supplemental material, sj-docx-4-nnr-10.1177_15459683251329893 for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study by Yu Wu, Depeng Zhou, Lovel Fornah, Jian Liu, Jun Zhao and Shicai Wu in Neurorehabilitation and Neural Repair

Supplemental Material

sj-jpg-1-nnr-10.1177_15459683251329893 – Supplemental material for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study

Supplemental material, sj-jpg-1-nnr-10.1177_15459683251329893 for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study by Yu Wu, Depeng Zhou, Lovel Fornah, Jian Liu, Jun Zhao and Shicai Wu in Neurorehabilitation and Neural Repair

Supplemental Material

sj-jpg-2-nnr-10.1177_15459683251329893 – Supplemental material for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study

Supplemental material, sj-jpg-2-nnr-10.1177_15459683251329893 for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study by Yu Wu, Depeng Zhou, Lovel Fornah, Jian Liu, Jun Zhao and Shicai Wu in Neurorehabilitation and Neural Repair

Supplemental Material

sj-png-3-nnr-10.1177_15459683251329893 – Supplemental material for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study

Supplemental material, sj-png-3-nnr-10.1177_15459683251329893 for Machine Learning-Based Model for Prediction of Early Post-Stroke Fatigue in Patients With Stroke: A Longitudinal Study by Yu Wu, Depeng Zhou, Lovel Fornah, Jian Liu, Jun Zhao and Shicai Wu in Neurorehabilitation and Neural Repair

Footnotes

Acknowledgements

We would like to thank all the participants in this study.

Author Contributions

Yu Wu: Conceptualization; Investigation; Methodology; Writing—original draft; and Writing—review & editing. Depeng Zhou: Methodology; Software; and Writing—review & editing. Lovel Fornah: Writing—original draft andWriting—review & editing. Jian Liu: Conceptualization; Investigation; and Writing—original draft. Jun Zhao: Project administration; Supervision; and Writing—original draft. Shicai Wu: Project administration; Supervision; Writing—original draft; and Writing—review & editing.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study received financial support from the National Key R&D Program of China (Grant No.: 2022YFC3600300 and 2022YFC3600305) and National Key R&D Program of China (Grant No.: 2023YFC3605200). The funder had no role in study design, data collection and analysis, preparation of the manuscript, or decision to submit the manuscript for publication.

Supplementary material for this article is available on the Neurorehabilitation & Neural Repair website along with the online version of this article.

ORCID iD

Yu Wu

References

GBD 2019 Stroke Collaborators, Global, regional, and national burden of stroke and its risk factors, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Neurol. 2021;20(10):795-820. doi:10.1016/S1474-4422(21)00252-0

Liu

, et al. China Stroke Study Collaboration. Stroke in China: advances and challenges in epidemiology, prevention, and management. Lancet Neurol. 2019;18(4):394-405.

National Stroke Center. Download information on the World Stroke Day 2023 theme campaign. Accessed October 20, 2023. https://sinosc.chinasdc.cn/NewsInfo/News/NewsDetailWeb?Tid=2899

Kang

Guo

, et al., Effect of IS history on hospital fitness in elderly patients with HFrEF. Chin J Geriatr Heart Brain Ves Dis. 2023;25(2):139-142.

Hinkle

Becker

Kim

, et al. American Heart Association Council on Cardiovascular and Stroke Nursing and Stroke Council. Poststroke fatigue: emerging evidence and approaches to management: a scientific statement for healthcare professionals from the American Heart Association. Stroke. 2017;48(7):e159-e170.

Wang

Prevalence of Post-stroke Fatigue after Ischemic Stroke and Research for Its Influencing Factors. Zhengzhou University; 2017.

Paciaroni

Acciarresi

Poststroke fatigue. Stroke. 2019;50(7):1927-1933.

Kirchberger

Wallner

Linseisen

, et al. Factors associated with early and late post-stroke fatigue in patients with mild impairment. results from the stroke cohort study Augsburg. Front Neurol. 2022;13:852486.

Lerdal

Gay

CL.

Fatigue in the acute phase after first stroke predicts poorer physical health 18 months later. Neurology. 2013;81(18):1581-1587.

10.

Pedersen

Almkvist

Holmegaard

, et al. Fatigue 7 years post-stroke: predictors and correlated features. Acta Neurol Scand. 2022;146(3):295-303.

11.

Lin

Chen

JX.

Investigation of fatigue in cerebral infarction patients. Chin J Nurs Educ. 2016;13(2):124-127.

12.

Kuppuswamy

Clark

Sandhu

, et al. Post-stroke fatigue: a problem of altered corticomotor control? J Neurol Neurosurg Psychiatry. 2015;86(8):902-904.

13.

Chen

Xiao

, et al., Poststroke fatigue: risk factors and its effect on functional status and health-related quality of life. Int J Stroke. 2015;10(4):506-512.

14.

Zang

Cui

, et al. A scoping review of assessment tools for post-stroke fatigue. Chin J Nurs. 2023;58(1):46-54.

15.

Kou

Qiao

, et al. Management of post-stroke fatigue: best evidence summary. Military Nurs. 2023;40(4):84-87.

16.

Luzum

Thrane

Aam

, et al. A machine learning approach to predict post-stroke fatigue. the nor-COAST study. Arch Phys Med Rehabil. 2024;105(5):921-929.

17.

Cotter

Salah Khlif

Bird

, et al. Post-stroke fatigue is associated with resting state posterior hypoactivity and prefrontal hyperactivity. Int J Stroke. 2021;27:17474930211048323.

18.

Jolly

Zainurin

Mead

, et al. Neuroimaging correlates of post-stroke fatigue: a systematic review and meta-analysis. Int J Stroke. 2023;18(9):1051-1062.

19.

Kutlubaev

Duncan

Mead

GE.

Biological correlates of post-stroke fatigue: a systematic review. Acta Neurol Scand. 2012;125(4):219-27.

20.

Zhang

Cheng

Zhang

, et al. Related risk factors associated with post-stroke fatigue: a systematic review and meta-analysis. Neurol Sci. 2021;42(4):1463-1471.

21.

Duncan

Anderson

, et al. Exploratory Cohort Study of Associations between Serum C: reactive protein and fatigue after stroke. PLoS One. 2015;10(11):e0143784.

22.

Liu

Wang

, et al. Elevated plasma high-sensitivity C-reactive protein at admission predicts the occurrence of post-stroke fatigue at 6 months after ischaemic stroke. Eur J Neurol. 2020;27(10):2022-2030.

23.

Ren

, et al. Serum uric acid levels at admission could predict the chronic post-stroke fatigue. Front Nutr. 2022;9:850355.

24.

Kliem

Labberton

Barra

, et al. Self-reported cognitive and psychiatric symptoms at 3 months predict single-item measures of fatigue and daytime sleep 12 months after ischemic stroke. Front Neurol. 2022;13:944586.

25.

Cui

Xia

Xing

, et al. Fatigue status and its influencing factors among patients in the acute stage of acute ischemic stroke. Nurs J Chin PLA. 2020;37(11):9-12.

26.

Collins

Reitsma

Altman

, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350:g7594.

27.

Diagnosis of various cerebrovascular diseases. Chin J Neurol. 1996;(6):60-61.

28.

, et al. A cross-sectional study of the major risk factor at different levels of cognitive performance within Chinese-origin middle-aged and elderly individuals. J Affect Disord. 2024;349:377-383.

29.

Prevalence, Influencing Factors of Post Stroke Fatigue and Impact on Quality of Life in Patients With Ischemic Stroke. Shandong University. 2019.

30.

Vittinghoff

McCulloch

CE.

Relaxing the rule of ten events per variable in logistic and Cox regression. Am J Epidemiol. 2007;165(6):710-718.

31.

Yan

Liu

Wang

, et al. Construction and validation of machine learning algorithms to predict chronic post-surgical pain among patients undergoing total knee arthroplasty. Pain Manag Nurs. 2023;24(6):627-633.

32.

Engel

GL.

The need for a new medical model: a challenge for biomedicine. Science. 1977;196(4286):129-136.

33.

Fugl-Meyer

Jääskö

Leyman

, et al. The post-stroke hemiplegic patient. 1. a method for evaluation of physical performance. Scand J Rehabil Med. 1975;7(1):13-31.

34.

Hsieh

Sheu

Hsueh

, et al. Trunk control as an early predictor of comprehensive activities of daily living function in stroke patients. Stroke. 2002;33(11):2626-2630.

35.

Alemseged

Rocco

Arba

, et al. Posterior National Institutes of Health Stroke Scale improves prognostic accuracy in posterior circulation stroke. Stroke. 2022;53(4):1247-1255.

36.

Hou

Zhang

, et al. Reliability and validity of the Chinese version of the National Institutes of Health Stroke Scale. Chin J Phys Med Rehabil. 2012;34(5):372-374.

37.

Buysse

Reynolds

3rd Monk

, et al. The Pittsburgh Sleep Quality Index: a new instrument for psychiatric practice and research. Psychiatry Res. 1989;28(2):193-213.

38.

Liu

Tang

, et al. Reliability and validity of Pittsburgh Sleep Quality Index. Chin J Psychiatry. 1996;29:103-107.

39.

Hoang

Le LT

Huong Nguyen

Tang

, et al. Investigating the traditional medicine shoulder pain (Jian Tong) characteristics in patients with ischaemic stroke in the early rehabilitation phase. Heliyon. 2024;10(2):e24626.

40.

Pathak

Sharma

Jensen

MP.

The utility and validity of pain intensity rating scales for use in developing countries. Pain Rep. 2018;3(5):e672.

41.

Liu

Qiao

, et al. The efficacy of computerized cognitive behavioral therapy for depressive and anxiety symptoms in patients with COVID-19: randomized controlled trial. J Med Internet Res. 2021;23(5):e26883.

42.

Yang

Wei

Zhang

, et al. Study on the correlation between thyroid nodules and anxiety and depression symptoms. psy. 2024;19(6):31-33.

43.

Hamilton

A rating scale for depression. J Neurol Neurosurg Psychiatry. 1960;23(1):56-62.

44.

Xiao

SY.

Theoretical basis and research application of Social Support Rating Scale. J Clin Psychiatry. 1994;4:98-100.

45.

Zeng

Guo

, et al. Latent profile analysis of financial toxicity and its influencing factors among stroke patients. Military Nurs. 2024;41(4):51-55.

46.

Smilkstein

Ashworth

Montano

Validity and reliability of the family APGAR as a test of family function. J Fam Pract. 1982;15(2):303-311.

47.

Family APGAR questionnaire and its clinical application. Foreign Med Sci. 1995;2:56-59.

48.

Krupp

LaRocca

Muir-Nash

, et al. The Fatigue Severity Scale. Application to patients with multiple sclerosis and systemic lupus erythematosus. Arch Neurol. 1989;46(10):1121-1123.

49.

Wang

DX.

Clinical application and assessment of the Chinese version of Fatigue Severity Scale in stroke patients. Chin J Phys Med Rehabil. 2007;29(9):608-611.

50.

Dablain

Krawczyk

Chawla

NV.

DeepSMOTE: fusing deep learning and SMOTE for imbalanced data. IEEE Trans Neural Netw Learn Syst. 2023;34(9):6390-6404.

51.

Priyadharshini

Banu

Sharma

, et al. Hybrid multi-label classification model for medical applications based on adaptive synthetic data and ensemble learning. Sensors (Basel). 2023;23(15):6836.

52.

Zhou

Wen

, et al. Using machine learning to build an early warning model for the risk of severe airflow limitation in patients with chronic obstructive pulmonary disease. Chin Gen Prac. 2022;25(2):217-226.

53.

Nahavandi

Alizadehsani

Khosravi

, et al. Application of artificial intelligence in wearable devices: opportunities and challenges. Comput Methods Programs Biomed. 2022;213:106541.

54.

Mato-Abad

Labiano-Fontcuberta

Rodríguez-Yáñez

, et al. Classification of radiologically isolated syndrome and clinically isolated syndrome with machine-learning techniques. Eur J Neurol. 2019;26(7):1000-1005.

55.

Xie

Wang

Xing

, et al. Identification of kidney-yang deficiency syndrome in osteoporosis patients based on rule ensemble method of bagging combining LASSO regression. Chin J Exp Tradit Med Formulae. 2023;29(23):150-157.

56.

Schnitzer

Hansson

Samuelsson

, et al. Fatigue in stroke survivors: a 5-year follow-up of the Fall study of Gothenburg. J Neurol. 2023;270(10):4812-4819.

57.

Snaphaan

van der Werf

de Leeuw

FE.

Time course and risk factors of post-stroke fatigue: a prospective cohort study. Eur J Neurol. 2011;18(4):611-617.

58.

White

Gray

Magin

, et al. Exploring the experience of post-stroke fatigue in community dwelling stroke survivors: a prospective qualitative study. Disabil Rehabil. 2012;34(16):1376-1384.

59.

Glader

Stegmayr

Asplund

Poststroke fatigue: a 2-year follow-up study of stroke patients in Sweden. Stroke. 2002;33(5):1327-1333.

60.

Guan

Gong

Research progress of post-stroke fatigue. Chin J Gerontol. 2016;36(11):2825-2827.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.02 MB

0.07 MB

0.78 MB

0.11 MB