Effectiveness of rehabilitation dose heterogeneity for improving functional outcomes of patients with acute stroke: a nationwide observational study

Abstract

To examine heterogeneity in the impact of increased early rehabilitation intensity and to identify which ischemic stroke patients benefit most. We analysed 39,140 propensity score-matched ischemic stroke patients using Diagnosis Procedure Combination database. Patients who received ≥ 60 min of daily rehabilitation within the 14 hospital days were compared with those who received less, in terms of achieving a modified Rankin Scale (mRS) score of 0, 1, or 2 at 60 days. We employed a causal forest model to estimate individualized treatment effectiveness and assessed for heterogeneity in rehabilitation impact across subgroups. Patients in the highest-benefit quartile showed substantial functional improvement (mean Conditional Average Treatment Effect: +17.1%; 95% CI, 15.3–19.0%), characterized by severe baseline activities of daily living (ADL) impairment, age, fewer comorbidities, higher socioeconomic status, better consciousness, and higher rates of reperfusion therapy. Conversely, patients in the lowest-benefit quartile, typically older with moderate ADL impairment, high comorbidities, low socioeconomic status, and limited reperfusion therapy, showed minimal or negative benefit (-16.5%; 95% CI, -18.1% to -14.9%). Rehabilitation dosing impact varied substantially regarding functional ability, with patients having severe strokes, age, fewer comorbidities, and better baseline consciousness deriving the greatest benefit from increased rehabilitation dosing.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-32006-w.

Introduction

Worldwide guidelines for stroke rehabilitation recommend delivering a sufficient “dose” of therapy to optimise functional recovery^1–4. Additionally, a systematic review reported that high-intensity doses improve functional ability⁵. These randomized controlled studies have demonstrated the effectiveness of rehabilitation dose heterogeneity, which appears to be influenced by various factors such as the methods used during early rehabilitation interventions⁶.

Despite the growing body of evidence regarding the effectiveness of rehabilitation dosing, previous studies that investigated the early rehabilitation dose intensity had three limitations^5,6. First, those randomized controlled and cohort studies focused on the effectiveness of high-dose rehabilitation in the intensive care unit, which limited the generalizability of their findings. Second, the study population included not only patients with stroke but also patients with various other diseases. Third, although some studies have identified variability in the effectiveness of rehabilitation based on the type provided, they did not examine patient-related factors. Consequently, it remains unclear which subgroups of patients with stroke are most likely to benefit from higher doses of early rehabilitation.

Several variables influence stroke recovery. For example, older age is associated with slow or less complete functional improvement^7,8. Moreover, a high comorbidity burden, as measured by the Charlson Comorbidity Index (CCI), can complicate rehabilitation outcomes⁹. Furthermore, a low socioeconomic status (SES) deteriorates the health status and limits access to medical resources^10,11, and the baseline functional status, such as the activities of daily living (ADL) score, is a strong predictor of recovery. Despite these findings, whether these factors directly mediate the impact of increased doses of early rehabilitation remains unclear.

Traditional subgroup analyses often stratify multidimensional patient data into simplified categories, thereby potentially overlooking important associations. In contrast, recent advances in machine learning have enabled researchers to estimate heterogeneous treatment effects by analyzing complex, high-dimensional data without such oversights¹². This approach allows the identification of specific patient characteristics associated with a favorable response to more intensive therapies¹³. By leveraging these methods, it is possible to assess how individual patient profiles modulate the impact of higher rehabilitation doses¹⁴. Such insights are particularly valuable in resource-limited settings in which prioritizing additional therapy for patients who are most likely to benefit is essential. Furthermore, for patients who are unlikely to respond to high-dose rehabilitation, these findings may guide the consideration of alternative treatment strategies. Therefore, we aimed to identify which patients with stroke would benefit most from an increased rehabilitation dose during early rehabilitation to help guide efficient resource allocation.

Methods

The Institutional Review Board at the Institute of Science approved this study (M2000-788-38), which complied with the ethical guidelines for medical and health science research involving human participants in Japan and the Declaration of Helsinki. The need for informed consent was waived because we analysed secondary anonymous data.

Data source and study population

We used data from the Diagnosis Procedure Combination per diem payment system, which is a representative principal database of acute care hospitals in Japan¹⁵. This database provides disease-related information such as primary diagnoses, associated conditions, medical procedures, and medications^15,16.

Participants

We extracted data of inpatients with ischaemic stroke (International Classification of Diseases, 10th Revision [ICD-10], code I63) who received rehabilitation across 1,089 hospitals between 1 April 2018 and 31 March 2019. We included patients 18 years of age and older with a pre-hospitalisation modified Rankin scale (mRS) score of 0, 1, or 2, which indicated that they were able to care for themselves. Patients were eligible if they experienced a stroke within 3 days, if their Japan coma scale (JCS) score was 0 or 1 (indicating no consciousness disorder or impaired consciousness but with eyes open), and if rehabilitation was started within 3 days.

We also included patients who were hospitalized for ≥ 14 days, as our aim was to observe the impact of rehabilitation during the first 14 days of hospitalisation.

Participants were excluded according to the following criteria: death within 24 h; craniotomy or decompression surgery; hospital stay longer than 180 days; progressive neurological diseases (ICD-10: G00–G99); use of artificial respiration; use of catecholamines; infection on admission; ambulatory status at admission (ADL: Mobility > 0); and hospitals with a subacute ward.

Because the mRS score at admission was not available, patients with ADL: Mobility > 0 were excluded to reduce the ceiling effect in the functional outcome measure.

Hospitals with a subacute ward were excluded because the presence of such wards could influence patient condition and care pathways during hospitalisation.

Exposure ascertainment

The mean daily rehabilitation duration during hospitalisation was calculated as the total duration of rehabilitation divided by the number of rehabilitation sessions performed from admission to day 14. Patients were categorised into the following two groups: those who received ≥ 60 min/day of rehabilitation and those who received < 60 min/day of rehabilitation. Rehabilitation is a multidisciplinary approach that comprises physiotherapy, occupational therapy, and speech therapy standardised according to guidelines. Rehabilitation services for inpatients are covered by universal health insurance in Japan.

Outcomes

The primary outcome was the proportion of patients who achieved an mRS score of 0, 1, or 2 at 60 days and 2 weeks, thus indicating a good outcome¹⁷.

In the Diagnosis Procedure Combination (DPC) database, mRS scores were recorded only at discharge.

To estimate the 60-day outcomes, we first identified patients who were discharged within 60 ± 14 days and used these cases as a training dataset. Functional recovery after stroke is most pronounced within 60 to 90 days after onset, as shown in previous studies¹⁸. Therefore, we defined the outcome window as 60 ± 14 days to capture this active recovery phase while minimizing missing data.

Patients whose outcomes were measured outside this period were treated as missing values. We trained a random forest model using outcome data from patients who were discharged within 60 ± 14 days and used it to impute the missing outcomes. We used a random forest model based on the following variables: length of hospital stay, nursing care needs, mRS score at discharge, JCS score at discharge, discharge destination, and use of home medical care after discharge.

In this analysis, death (mRS = 6) was categorized as a poor outcome, and poor outcomes were defined as mRS scores of 3–6. Overall, 11,984 patients (22.5%) had missing outcome data and required imputation.

Other covariates

We selected covariates based on existing evidence and clinical expertise. Key demographic variables included age, sex, pre-hospitalisation mRS score, and body mass index. Body mass index was calculated as weight (kg) divided by height squared (m²) and categorised into five groups based on the modified World Health Organization classification (18). Lifestyle factors, such as admission path, smoking status, and healthcare utilisation, including the use of long-term care services, were also included.

Comorbidities were assessed using the CCI, which was calculated according to Quan’s protocol based on ICD-10 codes (19), and the Hospital Frailty Index (HFI), which identifies frailty and predicts adverse healthcare outcomes (20). HFI scores were derived from ICD-10 codes, assigned specific point values, and stratified into the following two levels: low risk (< 5 points) and intermediate risk (> 5 points). Additionally, the consciousness level was categorised using the JCS (15) and grouped as follows: 0, conscious (unchanged); 1–3, awake without stimuli (converted to 1); 10–30, awake with stimuli (converted to 2); and 100–300, coma (converted to 3).

We also included comorbidities coded according to the ICD-10, such as diabetes mellitus, hypertension, dyslipidaemia, coronary artery disease, heart failure, atrial fibrillation, and dementia. Other hospital and clinical factors included readmission, use of emergency transportation, emergency admission, admission to specialised units such as the stroke care unit, high care unit, or intensive care unit, classification as a high-functioning hospital, and administration of intravenous thrombolysis (IVT). Interventions considered as covariates included dialysis, artificial ventilation, mechanical thrombectomy (MT), and the use of medication, such as warfarin, direct oral anticoagulants, antiplatelet drugs, and edaravone, during hospitalisation.

We classified SES based on the Japanese health insurance provider codes, which were grouped as high, middle, low, and unknown. High SES codes generally apply to employees or civil servants and include those that cover older adults with higher-level insurance plans. Middle SES codes typically correspond to municipal or regional insurers and include a designated code for older adults. Low SES codes usually indicate insurance schemes with fewer resources or coverage options. Unknown codes did not match any of these categories.

Statistical analysis

We conducted one-to-one propensity score matching without replacement to match participants who received ≥ 60 min of daily rehabilitation with those who received < 60 min of daily rehabilitation at baseline after adjusting for the following potential confounders: age, body mass index, JCS score, emergency transportation, CCI score, MT, IVT, care unit, total ADLs at admission, HFI, dementia, diabetes mellitus, hypertension, dyslipidemia, heart failure, urgency, mRS score before admission, admission path, long-term care grade, smoking status, and high-functioning hospital. We used logistic regression to calculate the propensity scores to determine rehabilitation durations ≥ 60 min and a caliper of 0.2 standard deviations of the logit of the propensity score¹⁹.

After matching the propensity scores, a causal model was implemented to predict the effect of heterogeneous treatment on the rate of mRS scores 0–2. This causal forest model consisted of an ensemble of 2000 trees and used honest splitting to minimize overfitting^12,20,21. The data subsample was divided as follows: the first half was used to define the tree structure, and the second half was reserved for prediction.

To further enhance the robustness of our model, 10-fold cross-fitting was applied during the training process. This method ensured that each tree was built using data subsets independent of the predictions. The parameters of the causal forest model, including the minimum node size and fraction of variables considered at each split, were turned using 10-fold cross-validation. Model calibration was assessed by fitting the best linear regression of observed outcomes to predicted treatment effects²⁰. The c-for-benefit was calculated to evaluate the discrimination performance of the model²².

We used a causal forest model to identify the subpopulations most likely to benefit from increased rehabilitation dosages. We estimated the conditional average treatment effect (CATE) of each patient, which represented the individualized effect of an increased rehabilitation dose, according to the baseline characteristics. Using these estimates, we stratified the patients into subgroups based on those in the quintile of the predicted distribution, thereby allowing us to explore heterogeneity in treatment responses.

Using baseline factors, we compared the characteristics of four groups and identified potential effect modifiers—variables that may influence the magnitude of the treatment effect—by examining their variable importance scores using the causal forest model. Only variables with importance scores ≥ 0.25 were selected for further consideration.

To robustly estimate the average treatment effect within each subgroup, we applied an augmented inverse probability weighting approach, which is a doubly robust method that combines propensity score modeling (to adjust for treatment assignment bias) and outcome regression modeling, that yielded consistent estimates even if one of the models was inaccurately specified²³. Then, we performed random forest-based multiple imputations. The imputation process was sequential and iterative. First, missing values were individually imputed using all other variables as predictors in a random forest model²⁴. The values predicted by the model were used as substitutes to fill in the missing values of the target variable. After the initial round of imputation, the substitute values were repeatedly used to retrain the random forest models. This iterative refinement continued until the imputed values stabilized, thereby ensuring reliable and robust imputation results. To examine the potential influence of missing data, a complete-case sensitivity analysis was additionally conducted and used as the main analytical approach for comparison with the imputed AIPW results.

A two-sided p < 0.05 and a standardized mean difference (SMD) < 0.1 were considered statistically significant. An SMD represents the difference in means or proportions divided by the pooled standard deviation, where a larger value indicates greater imbalance between groups. All analyses were performed using R software (version 4.2.2).

Results

Table 1; Fig. 1, and S1 Table present the baseline characteristics of the overall (non-adjusted) population (n = 53,366) and propensity score-matched population (n = 39,140). The matched and non-matched populations were similar in terms of the mean age (75.60 ± 11.53 vs. 75.70 ± 11.38 years, respectively). However, compared to the non-adjusted population, the matched population had a slightly lower proportion of patients with preserved consciousness (JCS score 0: 39.4% vs. 44.0%), higher rate of total ADL dependence at admission (44.0% vs. 33.5%), and lower proportion of patients without comorbidities (CCI score, 0: 39.4% vs. 44.0%).

Table 1.

Patient characteristics.

n	All	< 60 min	≥ 60 min	SMD
	53,366	33,750	19,616
Age, mean (SD)	75.60 (11.53)	75.60 (11.56)	75.61 (11.49)	0.001
Sex, female, n (%)	21,125 (39.6)	13,569 (40.2)	7556 (38.5)	0.034
BMI, n (%)				0.089
≤ 18	8997 (16.9)	5934 (17.6)	3063 (15.6)
19–24	31,101 (58.3)	19,821 (58.7)	11,280 (57.5)
25–29	10,995 (20.6)	6539 (19.4)	4456 (22.7)
≥ 30	2273 (4.3)	1456 (4.3)	817 (4.2)
JCS score, n (%)	28,591 (53.6)	19,300 (57.2)	9291 (47.4)	0.198
CCI score, n (%)				0.185
0	21,017 (39.4)	12,247 (36.3)	8770 (44.7)
1	18,791 (35.2)	12,804 (37.9)	5987 (30.5)
2	13,558 (25.4)	8699 (25.8)	4859 (24.8)
HFI, n (%)	6055 (11.3)	4018 (11.9)	2037 (10.4)	0.048
DM, n (%)	15,085 (28.3)	9744 (28.9)	5341 (27.2)	0.037
Hypertension, n (%)	28,157 (52.8)	17,659 (52.3)	10,498 (53.5)	0.024
Dyslipidemia, n (%)	17,027 (31.9)	10,620 (31.5)	6407 (32.7)	0.026
HF, n (%)	3769 (7.1)	2643 (7.8)	1126 (5.7)	0.083
Dementia, n (%)	2250 (4.2)	1470 (4.4)	780 (4.0)	0.019
Smoking, n (%)	23,848 (44.7)	14,845 (44.0)	9003 (45.9)	0.038
mRS score, n (%)				0.032
0	31,732 (59.5)	20,219 (59.9)	11,513 (58.7)
1	12,558 (23.5)	7932 (23.5)	4626 (23.6)
2	9076 (17.0)	5599 (16.6)	3477 (17.7)
Admission path				0.034
Home	50,289 (94.2)	31,857 (94.4)	18,432 (94.0)
Hospital/clinic	1445 (2.7)	932 (2.8)	513 (2.6)
Facility	1632 (3.1)	961 (2.8)	671 (3.4)
MT, n (%)	2929 (5.5)	2207 (6.5)	722 (3.7)	0.13
IVT, n (%)	4757 (8.9)	3243 (9.6)	1514 (7.7)	0.067
CU admission, n (%)	14,550 (27.3)	9524 (28.2)	5026 (25.6)	0.059
ADL status at admission, mean (SD)				0.399
Independent	964 (1.8)	566 (1.7)	398 (2.0)
Mild dependence	8629 (16.2)	4460 (13.2)	4169 (21.3)
Moderate dependence	13,180 (24.7)	7068 (20.9)	6112 (31.2)
Severe dependence	7117 (13.3)	4629 (13.7)	2488 (12.7)
Total dependence	23,476 (44.0)	17,027 (50.5)	6449 (32.9)
SES, n (%)				0.033
High	16,100 (30.2)	10,108 (29.9)	5992 (30.5)
Low	61 (0.1)	44 (0.1)	17 (0.1)
Middle	36,408 (68.2)	23,134 (68.5)	13,274 (67.7)
Other	797 (1.5)	464 (1.4)	333 (1.7)
Warfarin, n (%)	3927 (7.4)	2836 (8.4)	1091 (5.6)	0.112
DOAC, n (%)	14,119 (26.5)	9465 (28.0)	4654 (23.7)	0.099
Antiplatelet drugs, n (%)	51,290 (96.1)	32,431 (96.1)	18,859 (96.1)	0.003
Edaravone, n (%)	37,483 (70.2)	23,521 (69.7)	13,962 (71.2)	0.033
High-functioning hospital, n (%)	3080 (5.8)	2469 (7.3)	611 (3.1)	0.19

ADL, activities of daily living; BMI, body mass index; CCI, Charlson Comorbidity Index; CU, care unit; DOAC, direct oral anticoagulant; HF, heart failure; HFI, hospital frailty index; IVT, intravenous thrombolysis; JCS, Japan Coma Scale; mRS, Modified Rankin scale; MT, mechanical thrombectomy; SD, standard deviation; SES, socioeconomic status; SMD, standardised mean difference.

Fig. 1.

Study flow chart.

S1 Table presents baseline characteristics after propensity score matching, which showed good balance between the patient groups who received < 60 min and ≥ 60 min of daily rehabilitation across all variables. The performance of the causal forest model is summarized in S2 Table. The average predicted treatment effect across all individuals was 1.46, with a standard error of 1.74 and a t-value of 0.84 (p = 0.20), indicating that the overall.

average impact was not statistically significant. However, the heterogeneity of the treatment effects was significant. The differential forest prediction, which quantified the variation in treatment effects across individuals, was estimated to be 1.07, with a very small standard error of 0.04. The discrimination performance of the model, which was evaluated using the area under the receiver-operating characteristic curve for the treatment effect classification, was modest (0.581; 95% CI: 0.576–0.586) and suggested moderate ability to distinguish between patients with different levels of treatment responsiveness (S2 Table).

Figure 2 shows the distribution of the individual treatment effects estimated using the causal forest. Table 2 presents the estimated CATE with augmented inverse probability weighting of receiving ≥ 60 min of daily rehabilitation during the first 14 days of hospitalisation stratified by quartiles of the predicted treatment effect. The mean CATE varied substantially across quartiles, indicating clear heterogeneity in the impact of early intensive rehabilitation. Patients in the lowest quartile (quartile 1 [Q1]) demonstrated limited functional gains from intensive rehabilitation, with a mean CATE of − 16.5% (95% confidence interval [CI]: −18.1% to − 14.9%). In quartile 2 (Q2), the estimated benefit remained modest (mean CATE: −3.8%; 95% CI: −5.4% to − 2.2%). In contrast, patients in quartile 3 (Q3) showed a moderate functional improvement (mean CATE: +4.2%; 95% CI: 2.5%–5.9%). The greatest improvement was observed in quartile 4 (Q4) (mean CATE: +17.1%; 95% CI: 15.3%–19.0%). The variation in treatment effects indicated that rehabilitation intensity did not benefit all patients equally. Patients in the highest-benefit quartile (Q4) had approximately 17.1% higher probability of achieving functional independence (mRS 0–2) compared with those receiving lower-dose rehabilitation, whereas those in the lowest quartile (Q1) had a 16.5% lower probability. This gradient demonstrates that the response to intensive rehabilitation varied widely across individuals.

Fig. 2.

Factors that contribute to the impact of rehabilitation heterogeneity.

Table 2.

Distribution of the impact of ≥ 60 min of daily rehabilitation for patients with mRS scores 0–2 determined using AIPW.

	Mean CATE	Lower 95% CI	Upper 95% CI
Q1, %	−16.5	−18.1	−14.9
Q2, %	−3.8	−5.4	−2.2
Q3, %	4.2	2.5	5.9
Q4, %	17.1	15.3	19.0

AIPW, augmented inverse probability weighting; CATE, conditional average treatment effect; CI, confidence interval; mRS, modified Rankin scale; Q, quartile; SMD, standardised mean difference.

Table 3; Fig. 3 present the background characteristics categorised into quartiles (Q1–Q4) based on importance values with a standardised mean difference more than 0.25 for their degree of improvement with ≥ 60 min of daily rehabilitation (Fig. 2).

Table 3.

Determinants of the impact of rehabilitation heterogeneity.

n	Q1 (Lowest)	Q2	Q3	Q4 (Highest)	SMD
	13,342	13,341	13,341	13,342
ADL status at admission, n (%)					1.756
Independent	304 (2.3)	390 (2.9)	269 (2.0)	1 (0.0)
Mild dependence	3445 (25.8)	2873 (21.5)	2282 (17.1)	29 (0.2)
Moderate dependence	6625 (49.7)	4995 (37.4)	1558 (11.7)	2 (0.0)
Severe dependence	2160 (16.2)	1945 (14.6)	2604 (19.5)	408 (3.1)
Total dependence	808 (6.1)	3138 (23.5)	6628 (49.7)	12,902 (96.7)
CCI score, n (%)					0.84
0	2855 (21.4)	5057 (37.9)	5980 (44.8)	7125 (53.4)
1	2635 (19.7)	4849 (36.3)	5101 (38.2)	6206 (46.5)
2	7852 (58.9)	3435 (25.7)	2260 (16.9)	11 (0.1)
Age, mean (SD)	79.28 (9.18)	75.80 (10.52)	71.99 (13.75)	76.11 (10.98)	0.341
IVT, n (%)	368 (2.8)	604 (4.5)	955 (7.2)	2830 (21.2)	0.321
MT, n (%)	125 (0.9)	211 (1.6)	765 (5.7)	1828 (13.7)	0.299
JCS score, n (%)	5866 (44.0)	6345 (47.6)	6920 (51.9)	9460 (70.9)	0.295
SES, n (%)					0.286
High	2512 (18.8)	4334 (32.5)	5605 (42.0)	3649 (27.3)
Middle	10,667 (80.0)	19 (0.1)	16 (0.1)	9507 (71.3)
Low	17 (0.1)	8784 (65.8)	7450 (55.8)	9 (0.1)
Other	146 (1.1)	204 (1.5)	270 (2.0)	177 (1.3)

ADL, activities of daily living; BMI, body mass index; CCI, Charlson Comorbidity Index; IVT, intravenous thrombolysis; JCS, Japan Coma Scale; MT, mechanical thrombectomy; Q, quartile; SD, standard deviation; SES, socioeconomic status; SMD, standardised mean difference.

Fig. 3.

Distribution of the impact of ≥ 60 min of daily rehabilitation.

Patients who benefited the most (Q4) typically presented with severe baseline disability, older age, fewer comorbidities, better consciousness levels, and greater use of reperfusion therapy. In contrast, those in Q1 were generally older, moderately disabled at admission, and had multiple comorbidities. This pattern suggests that the benefits of intensive rehabilitation are greatest in patients who are initially more dependent but have physiological reserve and fewer systemic complications. The values of the SMD in reported Fig. 3 quantify the magnitude of baseline differences between the high-effect (Q4) and low-effect (Q1) groups, for example the standardised difference in mean age between Q4 and Q1 was 0.341.

Patients in Q1 tended to be the oldest (mean age, 79.28 ± 9.18 years). Additionally, age gradually decreased through Q2 (75.03 ± 10.52 years) and Q3 (71.99 ± 13.75 years) before increasing slightly in Q4 (76.11 ± 10.98 years). The proportion of patients with a JCS score of 0, which indicated preserved consciousness, also increased across quartiles (Q1: 44.0%; Q2: 47.6%; Q3: 51.9%; Q4: 70.9%), suggesting that a better neurological status was associated with greater functional gains. A clear trend in comorbidities was observed. The percentage of patients with a CCI score of 2 decreased progressively across quartiles (Q1: 58.9%; Q2: 25.7%; Q3: 16.9%; Q4: 0.1%). Conversely, the proportion of patients with a CCI score of 0 increased across quartiles (Q1–Q4: 21.4%–53.4%), reflecting a healthier baseline status in Q4.

The rate of reperfusion therapy followed a similar trend. The use of IVT increased from 2.8% in Q1 to 4.5% in Q2, 7.2% in Q3, and 21.2% in Q4. Similarly, MT was rarely performed in Q1 (0.9%) and Q2 (1.6%), but its rate increased to 5.7% in Q3 and 13.7% in Q4.

A striking pattern was observed in the baseline ADL status. Although only 6.1% of patients in Q1 were completely dependent on admission, this proportion increased sharply to 23.5% in Q2, 49.7% in Q3, and 96.7% in Q4. Conversely, independence and mild dependence were more common in Q1 and Q2 than in Q4, indicating that patients with severe functional impairment at baseline are more likely to achieve substantial improvement.

SES also varied across quartiles. Those with a high SES increased from 18.8% in Q1 to 32.5% in Q2, peaked at 42.0% in Q3, and decreased to 27.3% in Q4. Those with a low SES were predominant across all quartiles, although slightly less so in Q1 (0.1%) and Q4 (0.1%) than in Q2 (65.8%) and Q3 (55.8%).

In the sensitivity analysis using complete-case data (AIPW model), the distribution of the estimated impact of ≥ 60 min of daily rehabilitation among patients with mRS scores 0–2 showed consistent trends across quartiles (S3 Table). The mean conditional average treatment effect (CATE) increased from − 10.5% in Q1 to 15.4% in Q4, indicating that higher predicted benefit groups consistently gained greater impacts.

Variables contributing to the heterogeneity of treatment impact (defined as standardized mean difference ≥ 0.25) differed slightly from the main analysis. In the sensitivity analysis, the factors included Barthel Index category, JCS, IVT, MT, Hypertension, SES, Age, and Dyslipidemia, whereas CCI, Diabetes Mellitus, lesion location, and pre-admission mRS—which had been included in the main analysis—were not selected (S4 Table).

Discussion

This is the first study to examine the heterogeneous impact of early rehabilitation for patients with acute stroke using a causal forest model, thus enabling estimation of individualised treatment effects. Despite the well-documented average effects of early rehabilitation, clinical decision-making requires a deeper understanding of which patients will benefit most. The current study addressed this gap by applying a causal forest model to estimate individualised treatment effects on patients receiving stroke care. Although previous studies have examined average effects, our approach allowed a more nuanced understanding of which patient subgroups will benefit most from increased rehabilitation intensity. The impact of early rehabilitation on disability improvement is modified by a range of interacting patient-specific factors—including ADL status at admission, comorbidity burden, age, receipt of reperfusion therapy, level of consciousness, and socioeconomic status—even when the rehabilitation dose is the same. In our heterogeneity analysis, the greatest treatment benefits were observed among patients with poorer baseline ADL status, lower consciousness level, receipt of reperfusion therapy, and higher socioeconomic status, whereas those with greater comorbidity burden tended to show smaller gains. These findings indicate that rehabilitation alone does not determine recovery. Instead, the patient’s baseline function, physiological reserve, and access to acute-phase interventions shape the magnitude of benefit. The heterogeneity captured by the causal forest model therefore reflects meaningful clinical variation.

These findings suggest that patients with greater initial disability may benefit more from increased rehabilitation doses. Importantly, our results further indicate that the impact of rehabilitation is enhanced when combined with reperfusion therapy for patients with fewer comorbidities. Previous cohort studies have reported that rehabilitation dosing is more effective for patients with relatively severe stroke^25,26. In contrast, randomised controlled trials that targeted broader stroke populations have shown limited or no impact of higher rehabilitation doses^5,27,28. These discrepancies may be attributable to differences in patient background characteristics that influence responsiveness to rehabilitation²⁹. Additionally, our study suggests that not only patient severity but also the use of interventions during the acute phase, such as reperfusion, may play synergistic roles in optimising the benefits of high-dose rehabilitation.

Baseline ADL status

The ADL status was the strongest determinant of heterogeneity and is the most important prognostic factor for functional independence after acute stroke; additionally, beginning rehabilitation with a better ability to perform ADLs is associated with a much higher chance of achieving independence³⁰. Furthermore, increasing the rehabilitation intensity severely affects the baseline ADL status²⁹. These results indicate that patients with a severe ADL status require rehabilitation with sufficient intensity³¹. Our study provides evidence that dosing is beneficial for reducing the stroke severity.

Comorbidity

Another factor that determines heterogeneity is comorbidity. Comorbidities are independent predictors of poor rehabilitation outcomes^9,32. The presence of a high number of comorbidities is correlated with smaller functional gains and a higher risk of poor outcomes. High CCI scores often portend longer hospital stays and a lower chance of regaining independence³². Furthermore, a stroke-specific comorbidity score predicts rehabilitation outcomes, and the presence of a high number of comorbidities is associated with worse functional outcomes³³. In this study, a higher number of comorbidities was associated with lower impact of rehabilitation dosing, similar to previous studies. Globally, the comorbidity incidence rate has been increasing as the population ages³⁴. Therefore, providing rehabilitation dosing and programs that consider multiple comorbidities may be desirable.

Age

Age is an important determinant of heterogeneity and is well known to influence functional outcomes after stroke; although older patients can achieve substantial recovery, they often present with lower baseline function and greater comorbidity burdens^35,36. In our study, the relationship between age and treatment effect appeared generally linear from Q1 to Q3, with younger patients demonstrating larger effects. However, the highest-effect group (Q4) did not follow this pattern. This deviation likely reflects the multivariable interaction structure captured by the causal forest model, in which age is closely intertwined with other factors such as baseline ADL status, CCI score, level of consciousness, and receipt of acute treatments (e.g., IVT, MT). These factors may outweigh the negative impact of chronological age and create a subgroup of older patients with substantial recovery potential.

This interpretation aligns with previous large cohort studies showing that short-term high-intensity inpatient rehabilitation can yield comparable functional improvements across age groups—<65 years, 65–80 years, and > 80 years—when appropriate intensity and clinical management are provided^37,38. Our findings reinforce the notion that age alone should not limit the provision of higher-dose rehabilitation. Even among older patients, increased rehabilitation dosing can meaningfully contribute to functional improvement.

Reperfusion therapy

Reperfusion therapy, MT, and IVT are important factors associated with the impact of rehabilitation dosing because they may induce neural recovery and early neuroplastic stimulation during rehabilitation³⁹. Patients treated with IVT improved faster, and their functional ability improved to a greater extent⁴⁰. MT can dramatically reduce stroke severity by restoring blood flow, and a significant portion of recovery often occurs during the acute phase^41,42. These findings suggest that reperfusion therapy provides higher baseline function and may lead to additional gains with rehabilitation.

Level of consciousness

The consciousness level is a heterogeneous factor associated with rehabilitation dosing. Our study showed that a high rate of consciousness indicated high rehabilitation impact for patients with mRS scores 0–2. Studies have suggested that patients who start rehabilitation with impaired consciousness may achieve better recovery of the ability to perform ADLs^43,44. Our study included patients who had a coma scale score of 0–1 and those who were awake but demonstrated reduced clarity during communication and awareness. A lower consciousness level in these patients indicated good rehabilitation dosing outcomes.

Socioeconomic status

The SES is an important factor associated with heterogeneous rehabilitation dosing. A previous systematic review and meta-analysis revealed that SES is consistently associated with poorer functional outcomes after stroke⁴⁵. The lowest income group had a 36% higher risk of unfavorable outcomes, and survivors of stroke with a high income and patients who performed manual labor or were unemployed were likely to experience worse recovery⁴⁵. In this study, patients with stroke with a disadvantageous SES achieved less functional gain with rehabilitation dosing. This suggested that rehabilitation providers should understand the social determinants of health and explore approaches to rehabilitation interventions during the subacute phase and community reintegration that promote equal opportunities for those with a low SES.

Sensitivity analysis

In the complete-case sensitivity analysis, outcome missingness was primarily observed among patients with prolonged hospitalisation, who tended to have more severe conditions and were less likely to achieve discharge mRS 0–2. Consequently, the analytic population became biased toward those with shorter hospital stays and better recovery potential. This selection led to an overrepresentation of patients with higher functional improvement and reduced heterogeneity related to baseline comorbidity burden. As a result, the CATE distribution shifted upward, and the variables contributing to treatment effect heterogeneity differed slightly from the main analysis. Nevertheless, the consistent pattern of increasing CATE across quartiles suggests that the overall direction of treatment benefit remained robust despite potential selection bias introduced by outcome missingness and prolonged hospitalisation.

Limitations

Our study had some limitations. First, although the causal forest model was adjusted for many known factors, we did not include all possible influences. Unmeasured confounding factors may have affected these results. Although causal forests allow flexible estimation of heterogeneous treatment effects, their reliability depends on the assumption of no unmeasured confounding factors and sufficient overlap across treatment groups. Second, we described rehabilitation using only the average number of minutes of therapy per day. We did not record the specific type of therapy, content of each session, skills of the therapist, or activity of the patient during the session. Third, we evaluated success solely based on the attainment of functional independence, which was defined as an mRS score of 0–2 at discharge. However, the impact of rehabilitation is inherently multidimensional and encompasses aspects such as quality of life, cognitive function, and caregiver burden. Therefore, incorporating various outcome indicators is essential to comprehensively assessing the broader impact of rehabilitation. Fourth, comorbidities were identified using administrative codes from the DPC database, which capture conditions related to the hospitalisation episode but may not include all chronic diseases. Therefore, some misclassification or underreporting of comorbidities is possible. Fifth, although we compared our findings with previous studies, the study designs of those reports were not always specified in detail. Consequently, discrepancies between our results and prior studies may partly reflect methodological rather than clinical heterogeneity. We acknowledge this as a limitation. Sixth, we lacked detailed imaging information regarding the stroke itself, such as the exact brain region involved and size of the hemorrhage. Future studies should gather richer clinical and imaging data, describe the therapy in more detail, and evaluate a broader range of outcomes.

Finally, although this study focused on functional independence defined by mRS 0–2, this dichotomous measure may not fully represent the broader value of rehabilitation, especially for elderly or frail patients. Rehabilitation can yield meaningful benefits beyond functional recovery, such as improving quality of life and reducing caregiver burden. Future studies should include multidimensional, patient-centered outcomes to better capture the comprehensive benefits of rehabilitation.

Conclusion

Rehabilitation dosing has heterogeneous impact in terms of functional ability. Patients with stroke who have low ADL scores at admission benefit from increased rehabilitation dosing. We focused on functional ability, but rehabilitation infers other outcomes such as patient quality of life and reduction of caregiver burden. The need for rehabilitation dosing should be carefully considered.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Takuaki Tani and Masashi Kanai, Masatoshi Kamada, Kazushige Ide conceptualized and designed the study. Takuaki Tani drafted the main manuscript text. Kiyohide Fushimi contributed to critical revision of the manuscript for important intellectual content and supervised the overall research process. All authors reviewed and approved the final version of the manuscript.

Funding

This work was supported by the Japan Pharmaceutical Manufacturers Association; a Grant-in-Aid for Research on Policy Planning and Evaluation from the Ministry of Health, Labor, and Welfare, Japan [grant number 20AA2003]; and a Grant-in-Aid for Young Scientists (B) from the Japan Society for the Promotion of Science [grant number 21K10299].

Data availability

Datasets supporting the conclusions of this study are available from the authors upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

This study was approved by the Institutional Review Board and complied with the Declaration of Helsinki standards.