Functional recovery trajectories in post-acute stroke care: rethinking the length of stay Authors: Chao-Hung Yu, Ya-Wen Lee, Chih-Hao Lin, Hui-Yi Yang, Ying-Lin Hsu, Mu-Chien Sun, Chun-Yi Lin, Hao-Chih Yin, Hsiao-Fen Sun, Hui-Chun Tu, Pi-Ju Hsiao, Chih-Ming Lin Published: 2026 ABSTRACT AbstractBackgroundPost-acute care (PAC) restores function after stroke, but the optimal length of stay (LOS) is undefined, particularly in high-intensity rehabilitation settings, causing practice variation and resource-allocation challenges.ObjectiveTo examine the association between PAC LOS and functional outcomes after stroke using propensity score-based adjustment for baseline functional severity and to characterise recovery trajectories across rehabilitation durations.MethodsIn a retrospective cohort of patients enrolled in Taiwan’s PAC stroke programme (January 2018–December 2023), functional outcomes were assessed using the modified Rankin scale (mRS) and Barthel Index (BI). Associations between prolonged LOS (>6 weeks) and outcomes were examined using multivariable logistic regression and an inverse probability of treatment weighting model incorporating baseline functional status into the propensity score. Exploratory analyses evaluated weekly LOS cut-offs and stroke subtypes.ResultsAmong 610 patients, prolonged PAC LOS was associated with higher odds of mRS improvement (OR 1.75, 95% CI 1.20 to 2.57), lower odds of mRS non-improvement (0.57, 0.39 to 0.84) and lower odds of discharge BI≤60 (0.37, 0.25 to 0.55). Lower odds of BI non-improvement were also observed but rested on very few events (penalised-likelihood (Firth) estimation) and warrant caution. The decline in mRS non-improvement continued with longer LOS and differed by subtype, persisting beyond 6 weeks in ischaemic stroke.ConclusionsAfter adjustment for baseline functional status, prolonged PAC LOS remained associated with more favourable outcomes, but this should not be interpreted as evidence that longer rehabilitation independently causes better recovery. LOS appears to reflect individual recovery trajectories, whose timing differed by stroke subtype. These findings argue against a single uniform duration threshold and support individualised, subtype-aware reassessment.Keywords: STROKE WHAT IS ALREADY KNOWN ON THIS TOPIC Post-acute rehabilitation improves functional recovery after stroke, but the optimal length of stay is undefined and prior estimates are confounded because length of stay is shaped by each patient’s baseline severity and recovery course. Post-acute rehabilitation improves functional recovery after stroke, but the optimal length of stay is undefined and prior estimates are confounded because length of stay is shaped by each patient’s baseline severity and recovery course. WHAT THIS STUDY ADDS After incorporating baseline functional status into propensity-score weighting, prolonged post-acute care length of stay remained associated with more favourable functional outcomes; the benefit did not stop at a fixed point but continued beyond 6 weeks and its timing differed by stroke subtype, persisting longest in ischaemic stroke. After incorporating baseline functional status into propensity-score weighting, prolonged post-acute care length of stay remained associated with more favourable functional outcomes; the benefit did not stop at a fixed point but continued beyond 6 weeks and its timing differed by stroke subtype, persisting longest in ischaemic stroke. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY These findings argue against applying a single uniform length-of-stay threshold and support individualised, subtype-aware reassessment guided by each patient’s recovery trajectory, while highlighting the need for prospective, severity-adjusted studies to inform post-acute rehabilitation policy. These findings argue against applying a single uniform length-of-stay threshold and support individualised, subtype-aware reassessment guided by each patient’s recovery trajectory, while highlighting the need for prospective, severity-adjusted studies to inform post-acute rehabilitation policy. Introduction Stroke remains a leading cause of death and long-term disability worldwide.1 While advances in acute-phase treatments have improved survival, many patients experience persistent neurological deficits that necessitate intensive rehabilitation to restore daily activities and mitigate long-term disability.2 3 To address these challenges, many healthcare systems have implemented structured post-acute care (PAC) programmes,4 which are proven to be effective in improving functional status.5–7 Taiwan’s national PAC programme, introduced in 2014, targets stable patients with rehabilitation potential, transferring them to designated facilities for intensive, interdisciplinary therapy. Structurally, Taiwan’s PAC system operates similarly to inpatient rehabilitation facilities in the USA, which deliver high-intensity interdisciplinary therapy rather than skilled nursing facilities, which primarily provide lower-intensity skilled nursing care.8 9 Unlike lower-intensity models, this programme emphasises high-intensity functional recovery, where patients typically undergo 2–5 rehabilitation sessions per day, totaling at least 3 hours of therapeutic intervention daily. This model provides an initial stay of 3–6 weeks that is extendable up to 12 weeks based on functional assessments,10 similar to integrated stroke rehabilitation pathways found globally.11 12 Despite the well-supported efficacy of PAC, a critical clinical dilemma persists: the optimal length of stay remains undefined. Clinicians face the challenge of balancing the potential for further recovery against the risk of diminishing returns, where prolonged hospitalisations may not yield proportional benefits.13 14 This uncertainty is compounded by individual patient factors that influence recovery trajectories.15 Furthermore, the burden of stroke and the challenge of resource allocation are not limited to specific regions but are pragmatic global problems requiring systemic solutions.16 Amidst the pressures of ageing populations and healthcare resource constraints—a reality in both Taiwan and Europe—this lack of evidence-based guidance leads to significant practice variation and suboptimal resource allocation.17 18 Although previous studies have examined the association between rehabilitation length of stay and functional outcomes after stroke, the findings have remained inconsistent because patients requiring longer rehabilitation often present with greater baseline functional impairment. Consequently, the relationship between rehabilitation duration and functional outcomes remains difficult to interpret without adequate adjustment for baseline functional severity. A systematic evaluation of the relationship between PAC length of stay and patient outcomes is urgently needed to refine care pathways.19 Therefore, this study aimed to investigate the association between PAC length of stay and functional outcomes after stroke using propensity score-based adjustment incorporating baseline functional status. We further explored recovery patterns across different rehabilitation durations to inform individualised reassessment during post-acute stroke rehabilitation. Methods Study design and participants We conducted a retrospective cohort study using routinely collected clinical data from the Changhua Christian Hospital (CCH) healthcare system in Taiwan. The study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement (Research Checklist).20 Adult patients (aged ≥18 years) with acute ischaemic or haemorrhagic stroke who were consecutively admitted to the CCH PAC rehabilitation programme between 1 January 2018 and 31 December 2023 were screened for eligibility. Eligible patients were required to (1) have a confirmed diagnosis of ischaemic or haemorrhagic stroke; (2) be transferred from an acute care hospital within the CCH healthcare system; (3) have complete admission and discharge functional assessments and (4) meet the eligibility criteria for Taiwan’s PAC programme, including a baseline modified Rankin scale (mRS) score of 3–4 indicating moderate to moderately severe disability with rehabilitation potential. Patients were excluded if they had incomplete data for key study variables, died during the PAC admission or required transfer back to acute care before completion of rehabilitation. For descriptive analyses, patients were categorised according to functional improvement at discharge (mRS improvement vs no improvement) and PAC length of stay (>6 weeks vs ≤6 weeks). These groupings were used for descriptive comparisons only; the primary association analyses accounted for baseline differences using inverse probability of treatment weighting (IPTW), as described below. PAC programme and discharge criteria All patients participated in Taiwan’s National Health Insurance (NHI) PAC stroke rehabilitation programme, a standardised, high-intensity interdisciplinary rehabilitation model. The programme provides coordinated physical, occupational and speech therapy for at least 3 hours per day, 5–6 days per week, delivered according to nationally defined eligibility and treatment standards. Discharge decisions followed the standardised regulations of the Taiwan NHI PAC programme rather than individual physician preference. Patients were discharged when one of the following predefined criteria was met: (1) achievement of maximal functional recovery (functional plateau) as determined by the multidisciplinary rehabilitation team; (2) no meaningful functional improvement across two consecutive scheduled assessments; (3) development of medical complications requiring transfer back to acute care or (4) completion of the maximum reimbursable PAC duration (12 weeks). Although these standardised criteria reduce non-clinical variation in rehabilitation length of stay, discharge timing remains influenced by individual recovery trajectories and functional progress. Data collection and outcomes Clinical and rehabilitation data were extracted retrospectively from the electronic health record system of the CCH healthcare system using a standardised data abstraction process. Demographic characteristics, vascular risk factors, stroke subtype, treatment history, rehabilitation characteristics and functional assessment results at both PAC admission and discharge were collected for analysis. The primary outcome was functional improvement in global disability, defined as a decrease of at least one point in the mRS between PAC admission and discharge. This outcome was selected because the mRS is the principal functional outcome measure used for evaluating recovery in stroke rehabilitation and is the standard assessment tool within Taiwan’s NHI PAC programme. The secondary outcome was non-improvement in the Barthel Index (BI), defined as no increase or a decline in the BI score during the PAC admission. Unlike the mRS, which primarily reflects global disability, the BI evaluates independence in activities of daily living and is more sensitive to incremental functional gains achieved during rehabilitation. We therefore included BI non-improvement as a complementary outcome to identify patients who failed to achieve or maintain meaningful functional independence despite continued rehabilitation. Statistical analysis Continuous and ordinal variables were assessed using distributional diagnostics, including visual inspection of histograms and Q-Q plots, evaluation of skewness, outlier assessment, and examination of floor and ceiling effects. Age was approximately normally distributed and was summarised as mean with SD. Length-of-stay variables and functional assessment scales, including the mRS, BI, Functional Oral Intake Scale (FOIS), Mini Nutritional Assessment (MNA), EuroQol-5D (EQ-5D) and Instrumental Activities of Daily Living (IADL), were summarised as median with IQR because these measures were skewed, bounded, ordinal or demonstrated floor or ceiling effects. Categorical variables were summarised as frequencies and percentages. Between-group comparisons were performed according to variable type and distribution. Univariable logistic regression models included PAC length-of-stay exposure as the sole explanatory variable. Multivariable logistic regression models adjusted for age, sex, obesity, smoking, intravenous thrombolysis (tPA), endovascular thrombectomy (EVT), stroke subtype, hypertension, diabetes mellitus, hyperlipidaemia, heart disease, previous stroke, uraemia and baseline functional status, represented by baseline mRS and baseline BI. These covariates were prespecified a priori based on their clinical relevance and potential roles as confounders of the association between rehabilitation length of stay and functional outcomes; they were not selected according to univariable statistical significance or an automated stepwise procedure. Because the models were intended to estimate adjusted associations rather than to develop a prediction model, the effect of adjustment was evaluated by comparing the exposure ORs and 95% CIs between the univariable and multivariable models. Welch’s t-test was used for approximately normally distributed continuous variables, whereas the Mann–Whitney U test was applied to skewed continuous variables, length-of-stay variables and ordinal functional scales. Categorical variables were compared using the χ2 test or Fisher’s exact test when expected cell counts were small. To explore the temporal relationship between rehabilitation length of stay and functional recovery, weekly length-of-stay cut-offs were evaluated using multivariable logistic regression models incorporating the prespecified covariate set described above. In addition, restricted cubic spline (RCS) regression, adjusted for the same covariates, was used to characterise potential non-linear associations between PAC length of stay and functional outcomes. For outcomes or cut-offs with very few events, penalised-likelihood (Firth) logistic regression was used to obtain stable estimates. The primary exposure for the IPTW analyses was prolonged PAC length of stay, defined as >42 days. To reduce confounding by indication, IPTW based on propensity scores was applied to compare patients with PAC length of stay >42 days and ≤42 days. The propensity score model included age, sex, obesity, smoking, tPA, EVT, stroke subtype, hypertension, diabetes mellitus, hyperlipidaemia, heart disease, previous stroke, uraemia and baseline functional status, represented by baseline mRS and baseline BI. These covariates were prespecified a priori based on their clinical relevance and potential roles as confounders of the association between rehabilitation length of stay and functional outcomes; they were not selected according to univariable statistical significance or an automated stepwise procedure. National Institutes of Health Stroke Scale (NIHSS) was not included because this variable was unavailable in the validated dataset. Stabilised IPTW weights were calculated from the estimated propensity scores. Covariate balance before and after weighting was assessed using standardised mean differences (SMDs), with an absolute SMD<0.10 considered indicative of adequate balance. Propensity score overlap, weight distributions, effective sample size and extreme weights were evaluated as model diagnostics. Propensity score estimation was performed using complete-case analysis for the variables included in each model; patients with missing data for any included covariate were excluded from the corresponding analysis. Weighted logistic regression models were used to estimate the association between prolonged PAC length of stay and each binary functional outcome. Unweighted multivariable logistic regression models using the same covariate sets were additionally performed as comparator analyses. Because BI non-improvement occurred infrequently among patients with prolonged PAC length of stay, this outcome was interpreted cautiously with consideration of potential sparse-event instability. All statistical analyses were performed using R V.4.4.3 (R Foundation for Statistical Computing, Vienna, Austria). Two-sided p values<0.05 were considered statistically significant. Results Baseline characteristics A total of 703 patients admitted to the PAC programme for stroke rehabilitation were screened for eligibility. After excluding 93 patients because of death during the PAC admission (n=38) or transfer back to acute care (n=55), 610 patients were included in the final analysis (figure 1). The study cohort comprised 394 (64.6%) patients with ischaemic stroke and 216 (35.4%) with haemorrhagic stroke. Overall, 310 patients (50.8%) achieved an improvement of at least one point on the mRS by PAC discharge. Distribution of post-acute care (PAC) length of stay (LOS) by functional outcome. Histograms with overlaid density curves showing the distribution of PAC LOS according to functional outcome at discharge. Functional improvement was defined as a decrease of at least one point on the modified Rankin scale (mRS). Patients with LOS<2 weeks were retained to capture early discharge patterns. The figure illustrates the heterogeneous distribution of rehabilitation duration according to functional recovery status. Baseline demographic and clinical characteristics according to mRS improvement status are summarised in table 1. Compared with patients who achieved mRS improvement, those without improvement were older, had a higher prevalence of heart disease and previous stroke, and demonstrated lower functional status at PAC admission. Detailed demographic and clinical characteristics stratified by absolute discharge functional status (mRS≤2 vs. >2) are provided in online supplemental table S1. This cohort served as the basis for all subsequent exploratory and adjusted analyses. Table 1Baseline characteristics by mRS improvement statusCharacteristicStatisticTotalN=610mRS improvedn=310mRS non-improvedn=300P valueAge, yearsMean (SD)64.85 (13.58)61.67 (13.62)68.14 (12.74)<0.001Acute care LOS, daysMedian (IQR)16 (12–23)14 (11–22)17 (12–24)0.004ICU LOS, daysMedian (IQR)4 (2–8)4 (2–7)5 (3–9)0.246PAC LOS, daysMedian (IQR)41 (26–67)42 (31–65.8)40 (22–68)0.025Baseline mRSMedian (IQR)4 (4–4)4 (4–4)4 (4–4)0.006Baseline Barthel IndexMedian (IQR)45 (35–55)45 (40–60)40 (30–50)<0.001Baseline FOISMedian (IQR)7 (6–7)7 (7–7)7 (5–7)<0.001Baseline MNAMedian (IQR)10 (9–12)10 (9–12)10 (8–11)0.007Baseline EQ-5DMedian (IQR)10 (10–12)10 (9–11)10 (10–12)<0.001Baseline IADLMedian (IQR)1 (0–2)2 (1–3)1 (0–2)<0.001Discharge mRSMedian (IQR)3 (3-4)3 (2–3)4 (4–4)<0.001Discharge Barthel IndexMedian (IQR)70 (60–85)85 (70–90)60 (50–70)<0.001Discharge FOISMedian (IQR)7 (7–7)7 (7–7)7 (6–7)<0.001Discharge MNAMedian (IQR)10 (9–12)11 (9–12)10 (9–12)<0.001Discharge EQ-5DMedian (IQR)9 (8–10)8 (7–10)10 (9–10)<0.001Discharge IADLMedian (IQR)3 (1–4)4 (2–5)2 (1–3)<0.001SexFemale235 (38.5%)113 (36.5%)122 (40.7%)0.285ObesityYes165 (27.0%)94 (30.3%)71 (23.7%)0.064SmokingYes170 (27.9%)94 (30.3%)76 (25.3%)0.169HypertensionYes512 (83.9%)253 (81.6%)259 (86.3%)0.112Diabetes mellitusYes218 (35.7%)113 (36.5%)105 (35.0%)0.708HyperlipidemiaYes367 (60.2%)177 (57.1%)190 (63.3%)0.192Heart diseaseYes124 (20.3%)48 (15.5%)76 (25.3%)0.003Previous strokeYes129 (21.1%)46 (14.8%)83 (27.7%)<0.001UraemiaYes3 (0.5%)0 (0.0%)3 (1.0%)0.118Stroke subtypeHaemorrhagic216 (35.4%)120 (38.7%)96 (32.0%)0.083tPAYes51 (8.4%)19 (6.1%)32 (10.7%)0.043EVTYes49 (8.0%)23 (7.4%)26 (8.7%)0.571Continuous variables are presented as mean (SD) or median (IQR) according to distribution. Categorical variables are presented as n (%). P values were obtained using Welch’s t-test, Mann–Whitney U test, χ2 test or Fisher’s exact test as appropriate.EQ-5D, EuroQol-5D; EVT, endovascular thrombectomy; FOIS, Functional Oral Intake Scale; IADL, Instrumental Activities of Daily Living; ICU, intensive care unit; LOS, length of stay; MNA, Mini-Nutritional Assessment; mRS, modified Rankin scale; PAC, post-acute care; tPA, tissue plasminogen activator. Baseline characteristics by mRS improvement status Continuous variables are presented as mean (SD) or median (IQR) according to distribution. Categorical variables are presented as n (%). P values were obtained using Welch’s t-test, Mann–Whitney U test, χ2 test or Fisher’s exact test as appropriate. EQ-5D, EuroQol-5D; EVT, endovascular thrombectomy; FOIS, Functional Oral Intake Scale; IADL, Instrumental Activities of Daily Living; ICU, intensive care unit; LOS, length of stay; MNA, Mini-Nutritional Assessment; mRS, modified Rankin scale; PAC, post-acute care; tPA, tissue plasminogen activator. Recovery trajectories according to PAC length of stay The distribution of PAC length of stay according to mRS improvement status is shown in figure 1. Patients who achieved mRS improvement were most frequently discharged after approximately 5–6 weeks of rehabilitation, whereas patients without mRS improvement demonstrated a broader and more right-skewed distribution, reflecting greater variability in rehabilitation duration. These descriptive distributions provide an overview of discharge patterns but do not account for differences in baseline patient characteristics. Exploratory logistic regression analyses using weekly PAC length-of-stay cut-offs showed that prolonged rehabilitation was associated with lower odds of mRS non-improvement across the earlier weekly thresholds, and this association remained statistically significant at the 6 week cut-off (table 2). Estimates at the longest cut-offs (eg, 12 weeks) were based on very few patients remaining in rehabilitation and were therefore unstable; these are reported with caution and were additionally estimated using penalised-likelihood methods. Because the reference-group composition changes at each weekly cut-off, the weekly estimates fluctuated somewhat across thresholds; the RCS analysis (figure 2) provided a smoother characterisation of the underlying recovery trajectory. Table 2Exploratory analysis: adjusted ORs for mRS non-improvement by weekly PAC length of stay cut-offsLength of stay at PAC centreTotal populationIschaemic strokeHaemorrhage strokeCut-off (weeks)OR (95% CI)P valueOR (95% CI)P valueOR (95% CI)P value2 0.35 (0.18 to 0.68)0.0020.39 (0.17 to 0.91)0.0290.21 (0.06 to 0.70)0.01130.44 (0.27 to 0.72)0.0010.50 (0.27 to 0.93)0.0280.29 (0.11 to 0.76)0.0124 0.45 (0.29 to 0.70)<0.0010.53 (0.30 to 0.92)0.0230.30 (0.13 to 0.69)0.00550.47 (0.31 to 0.70)<0.0010.47 (0.28 to 0.79)0.0040.43 (0.20 to 0.93)0.0326 0.53 (0.35 to 0.79)0.0020.50 (0.30 to 0.85)0.0090.67 (0.33 to 1.34)0.25570.56 (0.38 to 0.84)0.0050.54 (0.32 to 0.91)0.0200.69 (0.35 to 1.38)0.29680.66 (0.44 to 0.99)0.0430.64 (0.38 to 1.09)0.1030.75 (0.38 to 1.49)0.40690.71 (0.45 to 1.12)0.1390.59 (0.32 to 1.07)0.0831.17 (0.57 to 2.39)0.669100.65 (0.40 to 1.05)0.0750.50 (0.26 to 0.97)0.0401.11 (0.53 to 2.35)0.784110.66 (0.40 to 1.08)0.1000.48 (0.24 to 0.96)0.0371.20 (0.57 to 2.53)0.63012 0.84 (0.10 to 6.92)0.8710.66 (0.04 to 10.86)0.7720.98 (0.03 to 29.12)0.991 ORs were derived from multivariable logistic regression models adjusted for age, sex, obesity, smoking, tPA, EVT, stroke subtype, hypertension, diabetes mellitus, hyperlipidaemia, heart disease, previous stroke, uraemia and baseline functional status (baseline mRS and baseline BI). The reference group for each cut-off is patients with a PAC length of stay less than or equal to the specified week. This exploratory analysis was performed to characterise temporal recovery patterns rather than to establish an optimal rehabilitation duration; interpretation should consider that the reference group composition varies at each timepoint. Estimates at the 12-week cut-off were based on very few exposed patients (2–5 per subgroup) and are unreliable even after penalised-likelihood (Firth) correction; they should not be interpreted.EVT, endovascular thrombectomy; mRS, modified Rankin scale; PAC, post-acute care; tPA, tissue plasminogen activator. Exploratory analysis: adjusted ORs for mRS non-improvement by weekly PAC length of stay cut-offs ORs were derived from multivariable logistic regression models adjusted for age, sex, obesity, smoking, tPA, EVT, stroke subtype, hypertension, diabetes mellitus, hyperlipidaemia, heart disease, previous stroke, uraemia and baseline functional status (baseline mRS and baseline BI). The reference group for each cut-off is patients with a PAC length of stay less than or equal to the specified week. This exploratory analysis was performed to characterise temporal recovery patterns rather than to establish an optimal rehabilitation duration; interpretation should consider that the reference group composition varies at each timepoint. Estimates at the 12-week cut-off were based on very few exposed patients (2–5 per subgroup) and are unreliable even after penalised-likelihood (Firth) correction; they should not be interpreted. EVT, endovascular thrombectomy; mRS, modified Rankin scale; PAC, post-acute care; tPA, tissue plasminogen activator. Non-linear relationship between post-acute care (PAC) length of stay and probability of modified Rankin Scale (mRS) non-improvement. Adjusted probability of mRS non-improvement as a continuous function of PAC length of stay, modelled using a 4-knot restricted cubic spline. Solid lines represent the adjusted probability of mRS non-improvement for the total cohort and for the ischaemic and haemorrhagic stroke subgroups; shaded areas represent the corresponding 95% CIs. Models were adjusted for demographic characteristics, vascular risk factors, treatment variables and baseline functional status (baseline mRS and baseline Barthel Index); the dashed vertical line marks the 6-week cut-off. In the total cohort, the probability of non-improvement declined rapidly during the first weeks of rehabilitation and flattened at approximately 7 weeks. Among patients with ischaemic stroke, the probability continued to decline until approximately 11 weeks, whereas among patients with haemorrhagic stroke the estimates were imprecise, with wide CIs at longer durations, and the length-of-stay–outcome relationship could not be reliably characterised. To further characterise the relationship between rehabilitation duration and functional recovery, RCS regression analyses were performed (figure 2). The adjusted RCS models demonstrated a non-linear association between PAC length of stay and the probability of mRS non-improvement. In the overall cohort, the probability of non-improvement declined rapidly during the first weeks of rehabilitation and flattened at approximately 7 weeks. This aggregate pattern, however, masked substantial heterogeneity by stroke subtype. When stratified by stroke subtype, patients with ischaemic stroke—who comprised the majority of the cohort—showed a slower decline in the probability of mRS non-improvement that continued well beyond 6 weeks and did not plateau until approximately eleven weeks. Among patients with haemorrhagic stroke, few patients remained at the longer durations and the corresponding estimates were highly imprecise; the length-of-stay–outcome relationship in this subgroup could therefore not be reliably characterised. These analyses provided the empirical basis for the subsequent evaluation of prolonged PAC length of stay using IPTW-adjusted analyses. IPTW-adjusted associations between PAC length of stay and functional outcomes The IPTW analyses included 609 patients with complete data for all variables in the propensity score model. Unadjusted discharge functional outcomes according to PAC length of stay (>6 weeks vs ≤6 weeks) are presented in online supplemental table S1B. Notably, for the mRS-based outcomes the unadjusted associations were directionally opposite to the adjusted estimates, consistent with confounding by the greater baseline severity of patients with longer length of stay. Following IPTW, baseline covariates were well balanced between the two length-of-stay groups, with all absolute SMDs below 0.10 (table 3). Table 3Baseline characteristics before and after IPTWCharacteristicBefore IPTW LOS>42 daysBefore IPTW LOS≤42 daysSMD beforeAfter IPTW LOS>42 daysAfter IPTW LOS≤42 daysSMD afterAge63.1566.12−0.22264.1864.85−0.050Male sex159 (60.9%)215 (61.8%)−0.01862.6%61.6%0.019Obesity65 (24.9%)100 (28.7%)−0.08727.2%27.7%−0.010Smoking75 (28.7%)94 (27.0%)0.03828.2%26.6%0.035tPA18 (6.9%)33 (9.5%)−0.0949.9%8.7%0.040EVT25 (9.6%)23 (6.6%)0.1098.2%8.6%−0.012Ischaemic stroke155 (59.4%)238 (68.4%)−0.18867.1%66.5%0.013Hypertension224 (85.8%)288 (82.8%)0.08479.9%83.3%−0.087Diabetes mellitus83 (31.8%)135 (38.8%)−0.14738.2%35.9%0.047Hyperlipidaemia156 (59.8%)211 (60.6%)−0.01862.4%60.7%0.033Heart disease44 (16.9%)80 (23.0%)−0.15418.5%20.6%−0.053Previous stroke49 (18.8%)80 (23.0%)−0.10424.8%21.9%0.069Uraemia1 (0.4%)2 (0.6%)−0.0280.5%0.5%0.000Baseline mRS3.943.780.4943.833.85−0.042Baseline Barthel Index38.9148.88−0.67845.1944.470.044SMDs are presented before and after weighting; negative values indicate the direction of the difference between the two length-of-stay groups. An absolute SMD<0.10 was considered indicative of adequate covariate balance. Weighted percentages are presented after IPTW.EVT, endovascular thrombectomy; IPTW, inverse probability of treatment weighting; LOS, length of stay; mRS, modified Rankin scale; SMDs, standardised mean differences; tPA, tissue plasminogen activator. Baseline characteristics before and after IPTW SMDs are presented before and after weighting; negative values indicate the direction of the difference between the two length-of-stay groups. An absolute SMD<0.10 was considered indicative of adequate covariate balance. Weighted percentages are presented after IPTW. EVT, endovascular thrombectomy; IPTW, inverse probability of treatment weighting; LOS, length of stay; mRS, modified Rankin scale; SMDs, standardised mean differences; tPA, tissue plasminogen activator. Compared with the unadjusted models, multivariable adjustment strengthened the observed associations between prolonged PAC length of stay and most functional outcomes. The OR for mRS improvement increased from 1.35 (95% CI 0.98 to 1.87) to 1.70 (95% CI 1.14 to 2.54), while the OR for mRS non-improvement decreased from 0.74 (95% CI 0.54 to 1.02) to 0.59 (95% CI 0.39 to 0.87). For discharge BI≤60, the OR decreased from 0.65 (95% CI 0.46 to 0.94) to 0.26 (95% CI 0.16 to 0.42). These changes were consistent with confounding by differences in baseline clinical and functional status between the length-of-stay groups. After IPTW adjustment incorporating baseline functional status, prolonged PAC length of stay (>6 weeks) remained significantly associated with higher odds of mRS improvement (OR 1.75, 95% CI 1.20 to 2.57) and lower odds of mRS non-improvement (OR 0.57, 95% CI 0.39 to 0.84), as well as lower odds of discharge BI≤60 (OR 0.37, 95% CI 0.25 to 0.55). A lower odds of BI non-improvement was also observed; however, because this outcome occurred in very few patients with prolonged length of stay, the estimate was obtained using penalised-likelihood (Firth) regression and should be regarded as a sparse-event finding warranting cautious interpretation (table 4). Table 4Association between PAC length of stay and functional outcomesOutcomeSparse-event flagUnadjusted modelMultivariable adjusted modelIPTW modelmRS improvementNo1.35 (0.98 to 1.87), p=0.0641.70 (1.14 to 2.54), p=0.0091.75 (1.20 to 2.57), p=0.004mRS non-improvementNo0.74 (0.54 to 1.02), p=0.0640.59 (0.39 to 0.87), p=0.0090.57 (0.39 to 0.84), p=0.004BI non-improvementYes0.08 (0.03 to 0.28), p=<0.0010.11 (0.03 to 0.37), p<0.0010.04 (0.01 to 0.17); Firth, sparse eventDischarge BI≤60No0.65 (0.46 to 0.94), p=0.0200.26 (0.16 to 0.42), p<0.0010.37 (0.25 to 0.55), p<0.001Values in the IPTW model reflect the propensity-score model incorporating baseline functional status (baseline mRS and baseline BI). BI non-improvement occurred in very few patients with prolonged length of stay and was estimated using penalised-likelihood (Firth) regression; it should be interpreted as a sparse-event result and not compared on the same inferential footing as the other outcomes. The unadjusted column reflects the exposure–outcome association without covariate adjustment; the multivariable-adjusted and IPTW columns reflect the covariate-adjusted and inverse probability of treatment weighting models, respectively.BI, Barthel Index; IPTW, inverse probability of treatment weighting; mRS, modified Rankin scale; PAC, post-acute care. Association between PAC length of stay and functional outcomes Values in the IPTW model reflect the propensity-score model incorporating baseline functional status (baseline mRS and baseline BI). BI non-improvement occurred in very few patients with prolonged length of stay and was estimated using penalised-likelihood (Firth) regression; it should be interpreted as a sparse-event result and not compared on the same inferential footing as the other outcomes. The unadjusted column reflects the exposure–outcome association without covariate adjustment; the multivariable-adjusted and IPTW columns reflect the covariate-adjusted and inverse probability of treatment weighting models, respectively. BI, Barthel Index; IPTW, inverse probability of treatment weighting; mRS, modified Rankin scale; PAC, post-acute care. Discussion After adjustment for baseline functional status and other measured confounders using inverse probability of treatment weighting, prolonged PAC length of stay remained significantly associated with more favourable functional outcomes, including higher odds of mRS improvement and lower odds of mRS non-improvement. Importantly, these associations should not be interpreted as evidence that longer rehabilitation duration independently causes better recovery. Rather, rehabilitation length of stay appears to reflect patients’ individual recovery trajectories within the PAC programme, with longer stays representing the clinical response to ongoing rehabilitation needs and continued functional progression rather than the therapeutic effect of additional hospitalisation itself. Consistent with this interpretation, the exploratory analyses did not show that the benefit of continued rehabilitation ended abruptly at a fixed time point; instead, the probability of mRS non-improvement continued to decline with longer length of stay, with the pace and timing of this decline differing across stroke subtypes.21 22 The findings for the mRS and the BI highlight that functional recovery after stroke is multidimensional rather than adequately represented by a single outcome measure. The mRS primarily reflects global disability and broad neurological status, whereas the BI focuses on independence in activities of daily living, which may continue to improve despite stabilisation of overall disability. Although the association between prolonged length of stay and reduced BI non-improvement was based on very few events and must therefore be interpreted cautiously, the broader pattern suggests that different functional outcome measures capture distinct aspects of the recovery process and should be regarded as complementary rather than interchangeable indicators of rehabilitation outcomes.7 22 From a clinical perspective, these findings caution against applying a fixed rehabilitation duration as a universal discharge criterion. Rather than treating any single time point as a uniform stopping rule, decisions regarding continuation or discharge from PAC rehabilitation should be individualised and guided by longitudinal functional assessment, stroke subtype and each patient’s ongoing trajectory of recovery. For many patients—particularly those with ischaemic stroke—meaningful functional gains may continue beyond 6 weeks, whereas for others a plateau in measurable improvement may support transition to a subsequent phase of care. This trajectory-informed approach may allow rehabilitation resources to be allocated according to individual recovery potential rather than a predetermined schedule.2 22 This study has several strengths. By incorporating baseline functional status into the propensity score model, the IPTW analysis substantially improved adjustment for clinically important differences between patients with shorter and longer rehabilitation stays, thereby strengthening the validity of the observed associations. Nevertheless, several limitations should be acknowledged. First, this was a retrospective single-centre study, and residual confounding from unmeasured factors cannot be completely excluded despite rigorous adjustment. In particular, the NIHSS, an important indicator of initial stroke severity, was unavailable in the validated dataset and therefore could not be incorporated into the propensity score model. In addition, IPTW can balance only measured covariates and therefore does not eliminate residual confounding from unavailable, inaccurately measured, or time-varying factors. Its validity also depends on correct specification of the propensity score model and adequate overlap in the probability of prolonged length of stay between exposure groups. Although we evaluated propensity score overlap, weight distributions, effective sample size, extreme weights and post-weighting covariate balance, satisfactory diagnostics cannot rule out unmeasured confounding. Moreover, prolonged length of stay was defined according to the completed duration of rehabilitation, meaning that patients had to remain in the PAC programme long enough to enter the >42-day group. Baseline IPTW cannot fully account for this time-dependent selection or for changes in functional status that influenced subsequent discharge timing. The weighted estimates should therefore be interpreted as adjusted associations and not as causal effects of extending rehabilitation beyond 42 days. Second, patients who died or required transfer back to acute care before completing rehabilitation (n=93) were excluded because standardised discharge functional assessments were unavailable for these individuals. Because such patients are likely to represent those with poorer clinical trajectories, their exclusion may introduce selection bias and preferentially remove less favourable outcomes from the analytical cohort. Although competing-risk approaches would provide additional insight into this issue, the retrospective dataset was not designed to support such analyses, and we therefore refrained from applying models that could not be adequately validated. Accordingly, our findings should be interpreted as applying to patients who completed the post-acute rehabilitation programme and underwent standardised outcome assessment, rather than to the broader population of all patients entering PAC; outcomes among patients with early clinical deterioration or interruption of rehabilitation warrant dedicated investigation in future prospective studies. Third, the relatively small number of BI non-improvement events among patients with prolonged PAC length of stay produced statistical instability, and this finding—estimated using penalised-likelihood methods—should therefore be interpreted with particular caution. Finally, the study was conducted within Taiwan’s NHI PAC programme, and the findings should be generalised to other healthcare systems with caution. Our findings provide a framework for reinterpreting rehabilitation length of stay within post-acute stroke care. Rather than viewing rehabilitation duration as an intervention that directly determines functional outcomes, future studies should investigate how longitudinal recovery trajectories can be integrated into individualised rehabilitation planning and discharge decision-making. Prospective multicentre studies incorporating comprehensive measures of neurological impairment, functional recovery and patient-centred outcomes will be important to validate and extend these observations across different healthcare settings. Such work may ultimately support a more individualised, trajectory-informed approach to post-acute stroke rehabilitation.22 Conclusions After adjustment for baseline functional status and other measured confounders, prolonged PAC length of stay remained associated with more favourable functional outcomes. These findings should not be interpreted as evidence that longer rehabilitation duration independently causes better recovery; rather, rehabilitation length of stay appears to reflect patients’ individual recovery trajectories and the clinical response to ongoing rehabilitation needs. Because the timing of functional plateau differed across stroke subtypes—with continued improvement beyond 6 weeks among patients with ischaemic stroke—these results argue against a single uniform duration threshold and support the use of longitudinal, subtype-aware functional assessment to guide reassessment and discharge planning. Further prospective, multicentre studies are warranted to validate this trajectory-informed approach to post-acute stroke rehabilitation. Supplementary material Acknowledgements We thank Tim Stubbings for manuscript proofreading. Permission to acknowledge has been obtained from the named individual. Footnotes Funding: The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors. Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Data availability free text: The datasets analysed during the current study are available from the corresponding author on reasonable request. Ethics approval: This study involves human participants and was approved by the Changhua Christian Hospital Institutional Review Board (IRB No. 241005). The study was granted a waiver of informed consent due to its retrospective nature and the use of de-identified data. Data availability statement Data are available upon reasonable request. References Associated Data Supplementary Materials Data Availability Statement Data are available upon reasonable request.