Bayesian Joint Longitudinal-Survival Modeling of Functional Recovery Trajectories and Time to Independent Community Ambulation Following Robotic Exoskeleton-Assisted Stroke Rehabilitation: A Multi-Centre Cohort Study in Canada

This multi-centre Canadian cohort study utilized a Bayesian joint modelling framework to demonstrate that both the current level and the rate of improvement in lower-extremity motor function are independently predictive of achieving independent community ambulation following robotic exoskeleton-assisted stroke rehabilitation, thereby supporting dynamic, trajectory-based treatment planning.

The Gist

Imagine you are watching a marathon, but instead of runners, the participants are stroke survivors trying to regain their ability to walk independently in their neighborhoods. For years, doctors have been trying to figure out two things: How fast are they getting better? and When will they cross the finish line?

This paper is like a high-tech, crystal-ball study that tries to answer those questions simultaneously using a special kind of math called a "Bayesian Joint Model."

Here is the story of the study, broken down into simple concepts:

1. The Problem: The "Two Separate Maps" Mistake

In the past, researchers looked at these two questions separately.

• Map A: They tracked how much a patient's leg strength improved week by week.
• Map B: They tracked how long it took for a patient to start walking outside alone.

The Flaw: This is like trying to predict when a car will reach a gas station by looking at the speedometer and the fuel gauge separately. In reality, they are connected! If a car is speeding up (getting stronger), it's more likely to reach the station sooner. If you ignore that connection, your predictions are often wrong.

2. The Solution: The "Super-Connected" Model

The researchers in this study built a new "Super-Connected" model. Instead of two separate maps, they built one giant, dynamic GPS system.

• The GPS: It tracks the patient's leg strength (measured by a test called the FMA-LE) as it changes over time.
• The Destination: It tracks the moment they achieve "Independent Community Ambulation" (walking outside without help).
• The Magic Link: The model realizes that how fast a patient is improving right now is just as important as how strong they are right now.

3. What They Discovered (The Plot Twist)

They followed 327 stroke survivors across Canada who used robotic exoskeletons (walking suits that help move legs) for rehabilitation. Here is what their "GPS" told them:

• The "Sprint" Phase: Recovery isn't a straight line. It's like a sprint followed by a jog. Most of the improvement happens in the first 12 weeks. After that, the gains slow down and eventually plateau (like a runner hitting a wall).
• The "Speed" Matters: This is the big discovery. Two patients might have the exact same leg strength score today.
  • Patient A has been improving slowly.
  • Patient B has been improving rapidly.
  • The Model says: Patient B is much more likely to start walking outside soon. The speed of their recovery is a secret predictor that doctors were missing before.
• The "Head Start" Advantage: Patients who started using the robotic suit sooner after their stroke had a much better chance of success. Waiting too long made the finish line harder to reach.

4. The "Crystal Ball" Feature

The most exciting part of this study is the Dynamic Prediction.

Imagine a doctor sitting with a patient at Week 4. Using this model, the doctor can say: "Based on how you improved in these first 4 weeks, here is your 70% chance of walking outside by Week 24."

Then, at Week 12, the doctor updates the prediction: "Great news, you improved faster than we thought! Your chance is now 85%."

It's like a weather forecast that updates every time a new cloud appears, giving a much more accurate picture of the future than a static guess made at the start.

5. Why This Matters for You

• For Patients: It gives hope and clarity. If you are improving fast, you are on a "fast track" to independence. If you are improving slowly, it doesn't mean you can't walk; it just means your doctor might need to change your training plan to help you speed up.
• For Doctors: It stops them from guessing. They can use the patient's current progress to decide if they need more robotic therapy, or if it's time to switch to a different type of training.
• For the System: Robotic exoskeletons are expensive. This model helps hospitals decide who will benefit the most, ensuring these high-tech tools are used where they work best.

The Bottom Line

This study is a breakthrough because it stopped treating "getting stronger" and "walking again" as two separate events. It showed they are two sides of the same coin. By using a smart, connected math model, the researchers proved that how fast you are getting better is a powerful crystal ball for predicting when you will be walking on your own.

It turns rehabilitation from a game of "wait and see" into a strategic game of "adjust and optimize."

Technical Summary

1. Problem Statement

Stroke rehabilitation research faces a critical methodological challenge: the interdependence between longitudinal functional recovery (e.g., motor score improvements over time) and time-to-event outcomes (e.g., achieving independent community ambulation).

• The Flaw in Standard Approaches: Traditional studies often analyze these outcomes separately. Longitudinal mixed-effects models assume "missing at random" (non-informative dropout), while survival models treat covariates as fixed or exogenous.
• The Reality: In rehabilitation, dropout is often informative (patients who recover faster may stop therapy earlier or achieve milestones sooner), and functional scores are endogenous (they evolve based on the patient's underlying recovery process). Separate analyses yield biased estimates and fail to capture the dynamic nature of recovery.
• The Gap: While Robotic Exoskeleton-Assisted Gait Training (RAGT) is a promising modality, its efficacy is often measured via static pre-post comparisons. There is a lack of evidence characterizing the nonlinear trajectory of recovery and its specific association with the probability of achieving clinically meaningful milestones like community ambulation.

2. Methodology

The study employed a Bayesian Shared-Parameter Joint Modeling framework to simultaneously estimate the longitudinal trajectory of motor function and the hazard of achieving community ambulation.

• Study Design & Population:

  • Type: Multi-centre retrospective cohort study (2019–2024) across four Canadian rehabilitation hospitals.
  • Sample: 327 adults with first-ever ischemic or hemorrhagic stroke receiving RAGT (EksoNR or Indego platforms).
  • Inclusion: Minimum 12 RAGT sessions; at least two post-baseline assessments.

• Outcome Measures:

  • Longitudinal Outcome: Fugl-Meyer Assessment Lower Extremity (FMA-LE) scores measured at baseline, weeks 4, 8, 12, 24, and 52.
  • Survival Outcome: Time to achieving Independent Community Ambulation, defined as a Functional Ambulation Category (FAC) score $\ge$ 4 sustained for two consecutive assessments (min. 4 weeks apart).

• Statistical Model Specification:

  • Longitudinal Submodel: A nonlinear mixed-effects model using a three-knot restricted cubic spline to capture the recovery curve.
    • $y_i(t) = m_i(t) + \epsilon_i(t)$
    • $m_i(t)$ includes fixed effects (time splines, covariates) and random effects (intercept, slope).
  • Survival Submodel: A Weibull proportional hazards model.
    • $h_i(t) = h_0(t) \exp(\gamma'w_i + \alpha_1 m_i(t) + \alpha_2 m_i'(t))$
    • Linking Mechanism: The hazard is linked to the subject-specific trajectory via two association parameters:
      • $\alpha_1$ (Current Value): Effect of the current FMA-LE level.
      • $\alpha_2$ (Slope): Effect of the rate of change (velocity) of the FMA-LE trajectory.
  • Estimation: Hamiltonian Monte Carlo (HMC) sampling via Stan (R package rstanarm).
    • 4 chains, 5,000 iterations (2,500 warmup).
    • Weakly informative priors specified for all parameters.
    • Convergence assessed via Gelman-Rubin statistic ($\hat{R} < 1.05$) and Effective Sample Size (ESS).

• Validation:

  • Model comparison using Leave-One-Out Information Criterion (LOOIC).
  • Dynamic prediction performance assessed via time-dependent AUC and Brier scores at 4, 12, and 24-week landmarks.
  • Five sensitivity analyses (e.g., alternative priors, current-value-only models, device stratification).

3. Key Contributions

1. Methodological Innovation: First application of a Bayesian joint model in robotic stroke rehabilitation literature to link nonlinear motor recovery trajectories with time-to-ambulation milestones, explicitly accounting for informative dropout and endogeneity.
2. Nonlinear Trajectory Characterization: Demonstrated that FMA-LE recovery follows a distinct nonlinear pattern (rapid initial gain followed by a plateau) rather than a linear progression, which standard linear models would misrepresent.
3. Dual Predictive Factors: Established that both the current functional level and the rate of change (slope) are independent predictors of achieving community ambulation.
4. Dynamic Prognostication: Developed a framework for real-time dynamic prediction, showing that prognostic accuracy improves significantly as more longitudinal data (FMA-LE scores) accumulates over time.

4. Key Results

• Recovery Trajectory:

  • Recovery was best described by a restricted cubic spline.
  • Rapid Phase: Mean gain of 8.7 FMA-LE points in the first 12 weeks (95% CrI: 7.2–10.3).
  • Plateau Phase: Gains decelerated significantly after week 12, with a near-plateau between weeks 24 and 52.
  • Covariates: Older age, higher baseline NIHSS, and hemorrhagic stroke were associated with lower overall trajectories.

• Association with Ambulation (Joint Model):

  • Current Value ($\alpha_1$): Each 1-point increase in the subject-specific FMA-LE trajectory increased the instantaneous hazard of achieving community ambulation by 8.8% (HR = 1.088, 95% CrI: 1.063–1.115).
  • Slope ($\alpha_2$): A steeper recovery gradient provided an additional 4.4% increase in hazard per unit slope (HR = 1.044), independent of the current score.
  • Baseline Predictors: Hemorrhagic stroke (HR=0.68), older age (HR=0.81 per decade), and higher baseline NIHSS (HR=0.94) were associated with lower hazards of achieving the milestone. Earlier RAGT initiation and higher baseline Berg Balance Scale scores were protective.

• Prediction Performance:

  • The model showed excellent discrimination, with the Area Under the Curve (AUC) increasing as more data became available:
    • 4-week landmark predicting 12-week outcome: AUC 0.78.
    • 24-week landmark predicting 52-week outcome: AUC 0.87.
  • Sensitivity analyses confirmed the robustness of the association parameters across different priors and model specifications.

5. Significance and Implications

• Clinical Practice: The findings support personalized rehabilitation planning. Clinicians should not rely solely on a patient's current FMA-LE score; the velocity of recovery is equally critical.
  • Patients with steep positive slopes may benefit from sustained high-intensity RAGT.
  • Patients with plateauing trajectories may require a reassessment of the intervention strategy or alternative therapies.
• Resource Allocation: The dynamic prediction framework can serve as a decision-support tool to optimize the allocation of scarce robotic exoskeleton resources by identifying patients with the highest probability of achieving community ambulation.
• Methodological Template: This study provides a rigorous template for future rehabilitation research, moving beyond static "pre-post" comparisons to dynamic, trajectory-based analyses that better reflect the biological reality of neuroplasticity and recovery.
• Policy: The results reinforce the economic and clinical value of RAGT, particularly when initiated early (within the first few months post-stroke) and tailored to individual recovery trajectories.