Machine learning model for predicting hospital-acquired functional decline in older patients with postoperative cardiovascular surgery

Abstract

BACKGROUND

Hospital-acquired functional decline (HAFD) is a poor prognostic factor in older patients who have undergone cardiovascular surgery.

AIM

To develop a model to predict HAFD and to identify its associated factors.

METHODS

This retrospective observational study included 144 patients who underwent cardiovascular surgery between May 2019 and December 2023. HAFD was defined as a change in the preoperative and pre-discharge short physical performance battery score. Seven machine learning models were constructed, and their performance was evaluated using the area under the receiver operating characteristic curve (AUC) values. The models were further interpreted using SHapley Additive exPlanations (SHAP) values.

RESULTS

Among the 144 participants, 41 (28.5%) experienced HAFD. Of the 7 machine learning models, the extreme gradient boosting model (XGBoost) achieved the best performance, with an AUC of 0.87. SHAP analysis revealed that being female and having a slower preoperative walking speed markedly impacted HAFD occurrence.

CONCLUSION

We developed a high-accuracy model to predict HAFD in older patients who have undergone cardiovascular surgery and identified key associated factors, informing preoperative evaluations and interventions in clinical practice.

Key Words: Cardiac surgery; Machine learning; Hospital-acquired functional decline; Extreme gradient boosting model; Predictive modeling

Core Tip: Hospital-acquired functional decline (HAFD) is a critical yet underrecognized complication in older patients undergoing cardiovascular surgery. We developed and validated a machine learning–based prediction model for HAFD using preoperative clinical and physical function data. Among seven models, the extreme gradient boosting model demonstrated the highest predictive performance. SHapley Additive exPlanations analysis identified female sex and slower preoperative walking speed as key contributors to HAFD. This interpretable model may support early risk stratification and targeted preoperative interventions to prevent functional decline in clinical practice.

INTRODUCTION

Cardiac rehabilitation following postoperative cardiovascular surgery in older patients effectively improves physical function and reduces postoperative complications and hospital stays[1]. It has been actively implemented in the early stages of hospitalization in many hospitals, partly owing to reported improvements in long-term prognosis[2] and the growing emphasis on shorter hospital stays in acute care settings[3].

However, despite postoperative cardiac rehabilitation, some patients experience limited physical recovery and require prolonged hospital stays owing to complications or other factors. Many of these patients present with frailty and preoperative declines in walking speed[4,5], suggesting the importance of these factors in predicting the prognosis of postoperative cardiovascular surgery in this patient population.

Recently, hospital-acquired functional decline (HAFD) has gained attention as a new predictive index in postoperative cardiovascular surgery in older patients[6,7]. Furthermore, HAFD is an independent predictor of poor prognosis two years after surgery[8]. Therefore, preoperative frailty and postoperative HAFD may have predictive value. Although HAFD is associated with poor prognosis, no studies have developed models to predict HAFD and its related factors.

Therefore, in this study, we aimed to develop a model to predict HAFD in postoperative cardiovascular surgery in older patients and to investigate the factors associated with its occurrence.

MATERIALS AND METHODS

Subjects

The study included 144 of the 173 patients who underwent standby cardiac surgery (coronary artery bypass, valvular, or mixed surgery) at our cardiovascular surgery center between May 2019 and December 2023. A total of 29 patients were excluded because they met at least one of the following predefined exclusion criteria: Difficulty in walking at admission, inability to undergo preoperative physical function assessment, or missing data. This study was approved by the ethical review committee of the Oita Oka Hospital. Patients were excluded based on the following criteria: (1) Difficulty in walking at the time of admission; (2) Difficulty in undergoing preoperative evaluation; and (3) Missing data.

Methods

This study was designed as a retrospective, observational study. Data were collected retrospectively from patients’ medical records.

Progress of postoperative rehabilitation

Postoperative rehabilitation was conducted according to the attending physician’s instructions, based on the “Criteria for the Start of Postoperative Weaning after Cardiac Surgery” and the “Criteria for Judging the Exercise Load Test (Step-up Criteria)” outlined in the Guidelines for Rehabilitation in Cardiovascular Disease (revised edition, 2021)[1], published by the Japanese Circulation Society.

Rehabilitation sessions were conducted twice daily for 40-60 minutes per session under the direction of the attending physician. In the rehabilitation room, patients primarily performed aerobic exercise using a bicycle ergometer and resistance training with equipment such as therabands and weights, in accordance with established rehabilitation guidelines.

Definition of HAFD

HAFD was defined as a decrease of one or more points in the short physical performance battery (SPPB) score before surgery and before hospital discharge[9]. The SPPB is a comprehensive physical function assessment tool for older individuals, consisting of balance, gait, and standing tests[10]. It is a validated and reliable measure commonly used to evaluate physical function in older populations[11].

The minimal clinically important difference for the SPPB is 1 point[12]. Accordingly, in this study, HAFD was defined as a decrease of 1 or more points in the preoperative SPPB score prior to hospital discharge.

Survey items

Data on age, sex, body mass index (BMI), comorbidities, and laboratory findings, including left ventricular ejection fraction as well as levels of brain natriuretic peptide, creatinine, urea nitrogen, hemoglobin, and C-reactive protein, were obtained from electronic medical records. Preoperative physical function was measured from admission through the day before surgery. Surgical procedure data were collected from surgical records.

Model building

Data analysis was conducted in three steps. First, feature selection was performed using LASSO regression with 5-fold cross-validation to improve generalization. Second, we split the dataset into training and test data, and we constructed predictive models using the features selected in the first step. These models included a logistic regression model, support vector machine (SVM), random forest, decision tree, the extreme gradient boosting (XGBoost), category boosting (CatBoost), and adaptive boosting (AdaBoost). Third, for the best-performing models, SHapley additive exPlanations (SHAP) values were calculated to investigate the effects of individual variables on the predictions made by the models (Figure 1). All analyses were performed using Python 3.913 software.

Figure 1 SHapley Additive exPlanations analysis of best-performing models. SHapley Additive exPlanations values were calculated for the best-performing models to investigate the effects of individual variables on predictions. The analysis demonstrates how each variable contributes to the model’s predictions for hospital-acquired functional decline. HAFD: Hospital-acquired functional decline.

RESULTS

A total of 144 patients were enrolled in the study, of whom 41 (28.5%) developed HAFD (Figure 2). Table 1 presents a comparison of the preoperative clinical characteristics of the HAFD and non-HAFD groups. The HAFD group was significantly older than the non-HAFD group (79.1 ± 6.6 years vs 74.9 ± 9.2 years, P = 0.009). Additionally, the HAFD group had a significantly higher proportion of female patients (53.7% vs 32.0%, P = 0.026) and faster walking speed (1.2 ± 0.5 milliseconds vs 1.1 ± 0.3 milliseconds, P = 0.020) compared with the non-HAFD group. No significant differences were observed between the two groups in terms of BMI, surgery type, comorbidities (hypertension, dyslipidemia, stroke, diabetes, orthopedic disorders, chronic obstructive pulmonary disease), laboratory data (hemoglobin, estimated glomerular filtration rate, B-type natriuretic peptide, creatinine, blood urea nitrogen, C-reactive protein), cardiopulmonary function (ejection fraction, % vital capacity, forced expiratory volume 1.0%), or other physical performance measures (5-component score, SPPB scores, grip strength). Seven variables with high coefficients were screened using LASSO regression, with HAFD as the dependent variable: Gender, preoperative walking speed, SPPB orthostasis, BMI, age, SPPB balance, and surgical technique (Figure 3). Based on these factors, seven models were constructed, with the XGBoost model yielding the best results, achieving an area under the receiver operating characteristic curve (AUC) of 0.87 (Figure 4). The SHAP values were then calculated for the XGBoost model, which revealed that being female and having a slow preoperative walking speed were significant factors influencing the model’s predictions (Figure 5).

Figure 2 Study flowchart of hospital-acquired functional decline incidence among study participants. A total of 144 patients were enrolled in the study, with 41 (28.5%) developing hospital-acquired functional decline (HAFD). The remaining 103 patients were in the non-HAFD group. AdaBoost: Adaptive boosting model; CatBoost: Category boosting model; SVM: Support vector machine model; XGBoost: Extreme gradient boosting model.

Figure 3 Variables screened using LASSO regression for hospital-acquired functional decline. A and B: Seven variables with high coefficients are screened using LASSO regression (A), and binomial deviance (B). LASSO: Least absolute shrinkage and selection operator.

Figure 4 Predictive performance of the extreme gradient boosting model assessed using a receiver operating characteristic curve. A: SHapley Additive exPlanations (SHAP) dependence plot; B: SHAP beeswarm plot. The extreme gradient boosting model (XGBoost) model performed the best for predicting hospital-acquired functional decline, achieving an area under the receiver operating characteristic curve value of 0.87. The XGBoost model is compared with other models. SPPB: Short physical performance battery; BMI: Body mass index.

Figure 5 SHapley Additive exPlanations analysis of the extreme gradient boosting model. A and B: SHapley Additive exPlanations (SHAP) values were calculated for the extreme gradient boosting model model to reveal the influence of individual variables on predictions from the SHAP dependence plot (A), and SHAP beeswarm plot (B). Being female and having a slower preoperative walking speed were key factors affecting the model’s predictions of hospital-acquired functional decline. XGBoost: Extreme gradient boosting model; CatBoost: Category boosting model; AdaBoost: Adaptive boosting model; AUC: Area under the receiver operating characteristic curve; SVM: Support vector machine model; ROC: Receiver operating characteristic.

Table 1 Preoperative clinical characteristics of the patients, n (%)/mean ± SD.

	Overall (n = 144)	HAFD group (n = 41)	Non-HAFD group (n = 103)	P value
Age (years)	76.1 ± 8.7	79.1 ± 6.6	74.9 ± 9.2	0.009
Sex (female)	55 (38.2)	22 (53.7)	33 (32.0)	0.026
BMI (kg/m2)	22.9 ± 3.2	23.6 ± 3.3	22.7 ± 3.1	0.105
Surgery type				0.056
CABG	63 (43.8)	12 (29.3)	51 (49.5)
Valve replacement	73 (50.7)	25 (61.0)	48 (46.6)
Combined	8 (5.6)	4 (9.8)	4 (3.9)
Comorbidity
Hypertension	110 (76.4)	30 (73.2)	80 (77.7)	0.722
Dyslipidemia	91 (63.2)	26 (63.4)	65 (63.1)	1.000
Stroke	16 (11.1)	3 (7.3)	13 (12.6)	0.535
Diabetes	59 (41.0)	14 (34.1)	45 (43.7)	0.388
Orthopedic disorders	40 (27.8)	13 (31.7)	27 (26.2)	0.647
COPD	5 (3.5)	2 (4.9)	3 (2.9)	0.939
Laboratory data
BNP (pg/mL)	246.1 ± 285.4	294.0 ± 321.6	227.1 ± 268.9	0.205
Creatinine (mg/dL)	1.2 ± 1.3	1.2 ± 1.1	1.3 ± 1.4	0.623
eGFR (mL/minute/1.73 m2)	54.0 ± 20.2	52.4 ± 20.2	54.7 ± 20.3	0.528
BUN (mg/dL)	21.0 ± 10.0	20.9 ± 9.8	21.1 ± 10.1	0.917
Hemoglobin (g/dL)	12.7 ± 2.1	12.3 ± 2.0	12.9 ± 2.1	0.148
CRP (mg/dL)	0.6 ± 1.3	0.4 ± 0.8	0.6 ± 1.5	0.423
EF (%)	55.3 ± 13.1	54.0 ± 14.5	55.8 ± 12.5	0.471
%VC (%)	91.6 ± 12.2	92.8 ± 9.4	91.2 ± 13.2	0.465
FEV1.0% (%)	78.7 ± 9.2	78.3 ± 10.1	78.9 ± 8.8	0.728
Walking speed (milliseconds)	1.1 ± 0.4	1.2 ± 0.5	1.1 ± 0.3	0.020
5-CS (second)	10.9 ± 4.2	11.5 ± 3.5	10.7 ± 4.4	0.310
SPPB balance score	3.8 ± 0.7	3.6 ± 0.8	3.8 ± 0.6	0.121
SPPB standing score	3.1 ± 1.2	3.3 ± 1.1	3.0 ± 1.3	0.342
SPPB walking score	3.6 ± 0.7	3.6 ± 0.8	3.6 ± 0.7	0.844
Grip strength (kg)	25.2 ± 8.4	23.1 ± 7.3	26.0 ± 8.7	0.058

The significance of differences between the hospital-acquired functional decline (HAFD) and non-HAFD groups was analyzed using t-tests or the χ² test. BMI: Body mass index; BNP: B-type natriuretic peptide; BUN: Blood urea nitrogen; CABG: Coronary artery bypass grafting; COPD: Chronic obstructive pulmonary disease; CRP: C-reactive protein; eGFR: Estimated glomerular filtration rate; EF: Ejection fraction; FEV1.0%: Forced expiratory volume % in one second; SPPB: Short physical performance battery; VC: Vital capacity.

DISCUSSION

The purpose of this study was to develop a model to predict HAFD in patients who have undergone cardiovascular surgery and to reveal factors associated with HAFD.

Factors influencing the model in our study included sex and preoperative walking speed. Postoperative complications and in-hospital mortality tend to be higher in women than in men[13]. This may be related to a slower recovery during postoperative rehabilitation, which could contribute to a decline in physical function. Walking speed has long been considered a prognostic factor in postoperative cardiovascular surgery[4,14]. Morisawa et al[15] reported that the risk of poor prognosis was 12 times higher in patients with HAFD associated with preoperative gait slowing than in those without gait slowing. This suggests that preoperative gait speed may modify HAFD, highlighting the importance of monitoring walking speed during preoperative assessments.

Machine learning has become an increasingly valuable tool in the medical field, assisting in diagnosis and prognosis by analyzing, training, and modeling medical data[16,17]. In our study, we utilized the XGBoost model, which is a gradient-boosting decision tree machine learning algorithm. This approach creates new trees by learning from the residuals of previous trees, enhancing generalization performance and robustness compared to those achieved by conventional SVMs and Random Forests. The XGBoost model is considered to have superior generalization performance and robustness[18,19]. In this study, overfitting was assumed to be prevented, and model performance was maximized by training with XGBoost.

We successfully developed a model to predict HAFD with high accuracy, using preoperative personal and physical function data as well as the XGBoost machine-learning model. The accuracy and reliability of the model were validated using test data, and the variables contributing to the model were analyzed using SHAP[20,21], which explains the influence of each variable. Our model may be implemented in actual clinical practice, potentially contributing to improvements in accuracy and clinical quality, with updates to the model data daily.

However, this study has some limitations. First, the data used to construct this model were collected retrospectively. Second, the sample size was small. Although the sample size of this study was insufficient to construct a machine learning model, an AUC of 0.87 was achieved, demonstrating the potential of the model. Further accumulation of data should improve accuracy of the model and broaden the scope of its clinical applications.

CONCLUSION

In summary, we constructed multiple machine learning models to predict HAFD in older patients who underwent postoperative cardiovascular surgery. We found that the XGBoost model provided the best predictive performance, with gender and preoperative walking speed identified as key relevant factors.

ACKNOWLEDGEMENTS

We thank the staff of the Oita Oka Hospital for their cooperation in data collection.