Early clinical indicators for predicting discharge destination from the acute stroke ward: A retrospective observational study

Abstract

In Japan, the length of stay in acute care hospitals has decreased, resulting in earlier discharge of stroke patients after life-saving treatment. This trend limits opportunities for patients to practice activities of daily living and for clinicians to prepare appropriate discharge plans. Early prediction of discharge destination is therefore essential to support timely rehabilitation and discharge management. This study aimed to develop a decision tree model to predict discharge destination using data obtained within 3 days of hospitalization. A retrospective observational study was conducted on 150 acute stroke patients. Clinical and demographic characteristics, medical history, cognitive status, National Institutes of Health Stroke Scale (NIHSS), Brunnstrom recovery stage, and motor- and cognitive-functional independence measure (M-FIM, C-FIM) scores were collected. Participants were randomly divided into training (n = 106) and test (n = 44) datasets. Ninety-one patients were discharged home and 59 to other facilities. NIHSS, M-FIM, and C-FIM were identified as key predictors. Patients with NIHSS ≤ 3.5, or with NIHSS > 6.5 combined with M-FIM > 23.5 and C-FIM > 29.5, had a 100% probability of home discharge, whereas those with NIHSS > 6.5 and M-FIM ≤ 23.5 had only a 14.7% chance. The model achieved 86.4% accuracy, with sensitivity of 91.7% and specificity of 80.0%. These findings suggest that combining early neurological and functional assessments provides a reliable basis for anticipating discharge outcomes, thereby aiding rehabilitation planning and effective patient management in acute stroke care.

1. Introduction

In Japan, the length of stay in acute care hospitals has decreased, leading to the prompt discharge of stroke patients who have completed life-saving treatment to their homes, rehabilitation wards, or long-term care facilities.^[1,2] This reduction in hospitalization duration limits the opportunity for patients to practice activities of daily living (ADLs) and make informed decisions regarding appropriate care services before discharge. Predicting the discharge destination soon after admission to an acute care hospital can help facilitate effective rehabilitation planning.^[3,4]

Several factors have been reported to influence the discharge destination of acute stroke patients, including age,^[5–7] neurological severity,^[8,9] ADL,^[10,11] and social background.^[12–14] Studies suggest that discharge destination is determined by a combination of factors rather than a single factor alone.^[2,14,15] One study by Iokawa et al^[2] used decision tree analysis to examine how multiple factors interact to affect discharge outcomes. Their findings indicated that stroke patients meeting the following criteria within 1 week of admission were more likely to be discharged home: a functional independence measure (FIM) score > 35, a Brunnstrom recovery stage (BRS) of the lower limb > 5, and a comprehension FIM score > 5. Decision tree analysis is a machine learning technique that develops predictive model by identifying combinations of factors that influence outcomes.^[16,17] Predictive models should be simple and practical,^[18–20] and decision trees are advantageous due to their intuitive, chart-based representation, making them easy to interpret and apply in clinical settings.^[16] These models provide clear stratification and visualization of key explanatory variables, enhancing their usability in clinical practice.^[21,22] Given the trend toward shorter hospital stays, the ability to predict a patient’s discharge destination immediately after admission is essential for effective rehabilitation and discharge planning. Rehabilitation in acute care hospitals typically begins within 3 days of stroke onset.^[23,24]

Thus, conducting a functional assessment of acute stroke patients within 3 days of admission can effectively predict their discharge destination, making it clinically valuable to determine discharge outcomes within this timeframe. This study aimed to develop a clinically useful decision tree model for predicting discharge destination based on assessments performed within 3 days of admission.

2. Methods

2.1. Study design

This study was a retrospective observational study approved by the Institutional Review Board (approval no. 700). The study was conducted in accordance with the Declaration of Helsinki and the Ethical Guidelines for Medical Research Involving Human Subjects. Due to the retrospective nature of the study, an opt-out method was employed. Information about the study was publicly disclosed, and participants were provided the opportunity to refuse participation.

2.2. Participants

Between April 2021 and December 2022, 216 stroke patients were admitted to the acute care unit in Japan. Of these, 66 patients were excluded based on the following criteria: 35 patients were already residing in a facility other than their own home before admission, 22 experienced significant deterioration during hospitalization (including worsening stroke symptoms, recurrence, new onset, or complications), and 9 were discharged due to death. As a result, 150 patients (length of stay, 40.9 ± 25.7 days) met the inclusion criteria and were included in the analysis (Fig. 1).

Figure 1.

Flowchart outlining the subject selection process.

2.3. Variables

Based on previous studies,^{[2,5,6,8,10–12,15]} the following factors were assessed within 3 days of admission: age, sex, presence of cohabiting family members (living alone or with family), preadmission modified Rankin Scale (mRS), stroke type (lacunar infarction, atherothrombotic infarction, cardiogenic infarction, intracerebral hemorrhage, or subarachnoid hemorrhage), medical history (including cerebrovascular disease, heart disease, respiratory disease, diabetes, dyslipidemia, or renal disease), presence of cognitive dysfunction, BRS (for the upper extremity, hand, and lower extremity), Glasgow Coma Scale, National Institutes of Health Stroke Scale (NIHSS), motor-FIM (M-FIM), and cognitive-FIM (C-FIM). The discharge destination was categorized as either discharge to home or other.

2.4. Statistical analysis

To perform hold-out validation, participants were randomly divided into a training dataset (70%) or a test dataset (30%). A classification and regression tree model using the Gini index was developed based on the training data, with discharge outcome (home or other) as the dependent variable and age, preadmission mRS; Glasgow Coma Scale; BRS for the upper limb, hand, and lower limb; NIHSS; M-FIM; and C-FIM as independent variables.

To prevent overfitting, the maximum tree depth was set to 3, the minimum number of cases required for a split was 10, and the minimum number of cases after splitting was 5. The test dataset was then applied to the model, and performance metrics—including accuracy, sensitivity, specificity, and F-score—were calculated using a confusion matrix. The decision tree analysis settings and validation methods followed previous studies.^[2,25–27] All statistical analyses were conducted using SPSS Statistics (IBM Corp., Armonk ), version 27.

3. Results

Among the 150 participants, 91 were discharged home, while 59 were transferred to other facilities. The average age was 77.0 ± 11.9 years, with a nearly equal distribution of men (82) and women (68). Lacunar infarction was the most common diagnosis, affecting 62 participants, followed by atherothrombotic infarction and intracerebral hemorrhage, each occurring in 38 cases (Fig. 1).

Cerebrovascular disease was the most frequently reported medical history, followed by diabetes mellitus. Additionally, approximately one-third of the participants (41 individuals) exhibited some degree of cognitive dysfunction (Table 1).

Table 1.

Participant characteristics.

		All patients	Home	Other
		n = 150	n = 91	n = 59
Age (mean ± SD)		77.0 ± 11.9	76.0 ± 12.8	79.9 ± 10.0
Sex (n)	Man/woman	82/68	52/39	30/23
Pre-hospitalization information
Preadmission mRS (n)	0/1/2/3/4/5	85/30/20/8/4/3	59/16/10/2/2/2	26/14/10/6/2/1
Presence of cohabiting family members (n)	Alone/with family	19/131	9/82	10/49
Stroke type
Lacunar infarction (n)		62	45	17
Atherothrombotic infarction (n)		39	21	18
Cardiogenic infarction (n)		5	2	3
Intracerebral hemorrhage (n)		39	21	18
Subarachnoid hemorrhage (n)		5	2	3
Medical history
Cerebrovascular disease (n)		41	22	19
Heart disease (n)		16	12	4
Respiratory disease (n)		3	3	0
Diabetes (n)		21	14	7
Dyslipidemia (n)		4	2	2
Renal disease (n)		3	1	2
Assessment
GCS (n)	Light/medium/heavy	113/31/6	83/7/1	30/24/5
BRS (n)
Upper limb	Ⅰ/Ⅱ/Ⅲ/Ⅳ/Ⅴ/Ⅵ	6/17/6/28/21/72	1/2/1/17/13/57	5/15/5/11/8/15
Hand	Ⅰ/Ⅱ/Ⅲ/Ⅳ/Ⅴ/Ⅵ	10/11/4/27/31/67	1/0/1/17/18/54	9/11/3/10/13/13
Lower limb	Ⅰ/Ⅱ/Ⅲ/Ⅳ/Ⅴ/Ⅵ	5/16/5/14/44/64	2/0/1/7/27/54	3/16/4/9/17/10
Presence of cognitive dysfunction (n)		44	15	29
NIHSS mean ± SD		7.7 ± 7.6	3.9 ± 3.9	13.4 ± 8.4
M-FIM mean ± SD		38.7 ± 24.2	50.1 ± 23.0	21.0 ± 12.5
C-FIM mean ± SD		24.8 ± 9.7	28.7 ± 8.1	18.9 ± 8.9

BRS = Brunnstrom recovery stage, C-FIM = cognitive-functional independence measure, GCS = Glasgow Coma Scale, M-FIM = motor-functional independence measure, mRS = modified Rankin Scale, NIHSS = National Institutes of Health Stroke Scale.

Participants were randomly assigned to a training dataset of 106 (about 70%) and a test dataset of 44 (about 30%). The decision tree constructed from the training data identified the highest probability of home discharge (100%) under the following conditions: NIHSS ≤ 3.5 or NIHSS > 6.5 but with M-FIM > 23.5 and C-FIM > 29.5 (Fig. 2).

Figure 2.

Decision tree analysis model predicting home and nonhome discharges, developed using 106 (approximately 70%) of the training data. C-FIM = cognitive-functional independence measure, M-FIM = motor-functional independence measure, NIHSS = National Institutes of Health Stroke Scale.

The lowest home discharge rate (14.7%) was observed in patients with an NIHSS >6.5 and an M-FIM of 23.5 or less. The model developed using the training data was validated with the test dataset, achieving an accuracy of 86.4%, sensitivity of 91.7%, specificity of 80.0%, and an F-score of 0.88 (Table 2).

Table 2.

Model validation results.

	Predicted
	Home (n)	Other (n)
Actual
Home (n)	22	2
Other (n)	4	16

Observed outcomes from the validation dataset (n = 44).

4. Discussion

This study developed a decision tree model to predict discharge destination using multiple variables within the first 3 days of admission in acute stroke patients.

The model incorporating NIHSS, M-FIM, and C-FIM demonstrated strong predictive performance, with an accuracy of 86.4%, sensitivity of 91.7%, specificity of 80.0%, and an F-score of 0.88, indicating high reliability in predicting home discharge.

In this study, the NIHSS score within the first 3 days of admission was identified as a key factor in predicting discharge to home. Previous research has shown a strong association between the NIHSS and functional outcomes as well as discharge destination.^[9,28] Naganuma et al^[29] found that a NIHSS cutoff of ≤8 at admission was linked to achieving independence in ADLs in patients with middle cerebral artery infarction.

In the first layer of our decision tree, the NIHSS cutoff of 6.5 aligns closely with Naganuma findings, supporting its relevance. Furthermore, our study provides new evidence that an NIHSS score of ≤3.5 within the first 3 days of admission significantly increases the likelihood of discharge to home.

Moreover, our decision tree shows that even with an NIHSS score >6.5, patients have a high likelihood of being discharged home if their M-FIM exceeds 23.5 and C-FIM exceeds 29.5. Although an NIHSS score above 6.5 typically indicates a moderate to severe stroke,^[30] this study found that higher independence in daily living, as measured by M-FIM and C-FIM, significantly increases the likelihood of home discharge, even in patients with severe neurological impairments. Previous studies have already shown an association between M-FIM and C-FIM scores and home discharge,^[10,31] and this study further identifies the FIM thresholds for predicting home discharge (i.e., M-FIM > 23.5 and C-FIM > 29.5).

The NIHSS and M-FIM are key factors influencing discharge destination,^[11,29] and this study suggests that when both scores are low, it becomes much more difficult for patients to be discharged home.

The novelty of this study lies in the development of a model that predicts the likelihood of home discharge based on assessments conducted within 3 days of admission to an acute care hospital. The strength of this study is its validation of the model’s high predictive accuracy using test data that was separate from the training data. A previous study^[2] created a decision tree to predict discharge destination (home, rehabilitation hospital, or other) using data collected within 7 days of admission, reporting an accuracy of 61.5% on the training data. While direct comparisons with previous studies are challenging due to different dependent variables, our study achieved a high accuracy of 86.4% on the test data, despite using information gathered within just the first 3 days of admission. The decision tree developed in this study, which accurately predicts the likelihood of home discharge early in the acute hospital stay, could aid in guiding early rehabilitation interventions tailored to the patient’s expected discharge destination.

4.1. Limitations section

This study has several limitations. First, it was conducted at a single hospital, which limits the generalizability of the results. Second, some factors associated with home discharge, such as stroke type,^[12] muscle strength,^[32] the number of family members,^[4] comorbidities,^[31] and pre-stroke physical activity,^[31] were not sufficiently assessed in this study. Although medical history was included as one of the variables, detailed analyses of comorbidity severity or cumulative burden were not performed. Moreover, although preadmission mRS was included as a proxy for pre-stroke physical activity, detailed information on patients’ activity levels before onset (e.g., exercise habits or activity frequency) was not available. Third, the results of this study were derived from patients with relatively long hospital stays. The length of stay may affect discharge outcomes; therefore, the generalization of the model to settings with shorter hospitalizations should be interpreted with caution. Future studies should address these factors to provide further insights.

5. Conclusions

This study developed a simple and accurate predictive model to assess the likelihood of home discharge in acute stroke patients within 3 days of hospitalization. Based on the NIHSS and FIM scores, the model showed high accuracy and clinical utility, offering valuable insights for early rehabilitation planning and discharge support.

Acknowledgments

The authors would like to express their gratitude to all patients who participated in this study.

Author contributions

Conceptualization: Takaya Komiyama, Kenji Tsuchiya, Takaaki Fujita, Kohei Obuchi, Fusae Tozato.

Data curation: Takaya Komiyama.

Investigation: Takaya Komiyama, Kohei Obuchi.

Methodology: Takaya Komiyama, Kenji Tsuchiya, Takaaki Fujita, Fusae Tozato.

Project administration: Takaya Komiyama, Kenji Tsuchiya, Fusae Tozato.

Software: Takaya Komiyama.

Supervision: Kenji Tsuchiya, Fusae Tozato.

Writing – original draft: Takaya Komiyama, Kenji Tsuchiya, Takaaki Fujita, Fusae Tozato.

Writing – review & editing: Kenji Tsuchiya, Takaaki Fujita, Fusae Tozato.