Michael H. Schwartz

Aim: To estimate the effect of surgery on parent‐reported gait outcomes as measured using the Gait Outcomes Assessment List (GOAL) questionnaire in a cohort undergoing surgery guided by instrumented gait analysis. Method: This was an observational study ( n = 146 individuals, 59 females [40%], median age 7 years 10 months [interquartile range 6 years to 10 years 10 months]) classified in Gross Motor Function Classification System levels I–III). The X‐Learner was used to estimate the causal treatment effect on treated individuals. The X‐Learner is an advanced machine learning framework designed to estimate causal and heterogeneous treatment effects while controlling for confounding and minimizing bias. Changes in parent‐reported GOAL domains were compared with objective measures (Gait Deviation Index, walking speed, energy consumption); heterogeneous treatment effects were explored. Results: Significant positive causal treatment effects were observed in gait pattern and appearance, and use of braces and mobility aids (+10 points). Moderate effects were observed for body image and self‐esteem, and pain, discomfort, and fatigue (+5 points). Functional domains (activities of daily living and independence; gait function and mobility; physical activities, sports, and recreation) showed only small positive effects or slightly negative effects. Gait pattern and appearance was positively correlated with the Gait Deviation Index; however, GOAL items related to walking speed and fatigue showed no association with laboratory‐measured analogs. Interpretation: Surgery guided by instrumented gait analysis substantially improved gait appearance and comfort. Clinicians should manage expectations regarding improvements in functional mobility. The disconnect between subjective patient‐reported outcome measures and objective measures highlights the complexity of quantifying a treatment's full impact.

Aim: To test whether selective dorsal rhizotomy (SDR) changes the incidence of subsequent orthopedic surgery in children with cerebral palsy. Method: This was a matched cohort study. Using historical data from a single center, we identified two groups of individuals based on their history of SDR (yes‐SDR and no‐SDR), matched at baseline on key clinical variables. We defined 10 orthopedic surgery categories that accounted for more than 95% of surgery recorded in our gait center database. We then fitted Kaplan–Meier estimates of cumulative incidence for each surgery, stratified according to SDR status. Results: We obtained excellent comprehensive baseline matching. Foot and ankle bone surgery and femoral derotation osteotomy exhibited 15% to 20% higher cumulative incidence after SDR. Rectus femoris transfer exhibited approximately 30% lower cumulative incidence after SDR. The remaining surgeries, including all surgeries for treating muscle contractures, were unaffected by SDR, exhibiting only small ( 0.05, log‐rank test) differences in cumulative incidence. Interpretation: SDR does not consistently lower the cumulative incidence of orthopedic surgery and does not lower the rate of surgery for muscle contracture. The mechanisms through which SDR influences subsequent surgery remain unclear, highlighting a critical area for future research.

Clinical gait analysis (CGA) has historically relied on clinician experience and judgment, leading to modest, stagnant, and unpredictable outcomes. This paper introduces Evidence-Based Gait Analysis Interpretation Tools (EB-GAIT), a novel framework leveraging machine learning to support treatment decisions. The core of EB-GAIT consists of two key components: (1) treatment recommendation models, which are models that estimate the probability of specific surgeries based on historical standard-of-practice (SOP), and (2) treatment outcome models, which predict changes in patient characteristics following treatment or natural history. Using Bayesian Additive Regression Trees (BART), we developed and validated treatment recommendation models for 12 common surgeries that account for more than 95% of the surgery recorded in our CGA center’s database. These models demonstrated high balanced accuracy, sensitivity, and specificity. We used Shapley values for the models to enhances interpretability and allow clinicians and patients to understand the factors driving treatment recommendations. We also developed treatment outcome models for over 20 common outcome measures. These models were found to be unbiased, with reliable prediction intervals and accuracy comparable to experimental measurement error. We illustrated the application of EB-GAIT through a case study, showcasing its utility in providing treatment recommendations and outcome predictions. We then use simulations to show that combining recommendation and outcome models offers the possibility to improve outcomes for treated limbs, maintain outcomes for untreated limbs, and reduce the number of surgeries performed. For example, under the counterfactual situation where femoral derotation osteotomies are administered only when they align with historical standard of practice (> 50% probability of surgery) and are predicted to improve the Gait Deviation Index (change > 7.5 points), the model predicts a 11 percentage point reduction in surgeries (26% limbs currently, 15% limbs simulated), a 6 point improvement in Gait Deviation Index among treated limbs (6 currently, 12 simulated), and no change in Gait Deviation Index for untreated limbs (2 currently, 2 simulated). EB-GAIT represents a significant step toward precision medicine in CGA, offering a promising tool to enhance treatment outcomes and patient care. The EB-GAIT approach addresses the limitations of the conventional CGA interpretation method, offering a more structured and data-driven decision-making process. EB-GAIT is not intended to replace clinical judgment but to supplement it, providing clinicians with a second opinion grounded in historical data and predictive analytics. While the models perform well, their effectiveness is constrained by historical variability in treatment decisions and the inherent complexity of patient outcomes. Future efforts should focus on refining model inputs, incorporating surgical details, and pooling data from multiple centers to improve generalizability.

We used a Bayesian model to study oxygen kinetics in children with CP. First, we quantified the oxygen kinetics time constant, finding no significant differences across GMFCS, sex, or age despite slower kinetics compared with typically developing peers. Next, we validated the model’s ability to estimate steady-state V̇o₂ with only 3 min of data, halving the usual testing time. This method has important implications for faster, accurate metabolic analyses in clinical and research settings.

Background Pediatric traumatic brain injury can lead to severe disability. Currently used standard clinical measures effectively capture secondary functional impairments, but do not measure neurologic impairment directly. Objective Evaluate the feasibility of walking dynamic motor control (walk-DMC) assessments to more directly measure neurological impairment and recovery for individuals post-traumatic brain injury. Methods The trajectory of walk-DMC and standard clinical measures of balance, mobility, and function were assessed in a cohort of individuals post-traumatic brain injury. Measures were collected throughout participants’ inpatient rehabilitation stay and at short- and long-term follow-up assessments. Results Four pediatric participants with severe traumatic brain injury enrolled. All participants demonstrated substantial neurological impairment at enrollment. All clinical measures showed an initial deficit followed by recovery, with most returning to nondisabled ranges over the study period for all participants. In contrast, walk-DMC scores demonstrated an initial acute deficit and did not reach nondisabled ranges for two of the participants, indicating persistent neurologic impairment. Conclusion Walk-DMC shows promise in its ability to identify subtle ongoing neurologic impairment compared to traditionally used clinical assessments of balance, mobility, and function. Further work in a larger cohort of participants with traumatic brain injury will improve understanding of how walking dynamic motor control changes with injury severity and where such a measure can serve as a leading indicator of neurologic and functional recovery.

Clinical gait analysis (CGA) has historically relied on clinician experience and judgment, leading to modest, stagnant, and unpredictable outcomes. This paper introduces Evidence-Based Gait Analysis Interpretation Tools (EB-GAIT), a novel framework leveraging machine learning to support treatment decisions. The core of EB-GAIT consists of two key components: (1) treatment recommendation models , which are propensity models that estimate the probability of specific surgeries based on historical standard-of-practice (SOP), and (2) treatment outcome models , which predict changes in patient characteristics following treatment or natural history. Using Bayesian Additive Regression Trees (BART), we developed and validated treatment recommendation models for 12 common surgeries that account for more than 95% of the surgery recorded in our CGA center’s database. These models demonstrated high balanced accuracy, sensitivity, and specificity. We used Shapley values for the models to enhances interpretability and allow clinicians and patients to understand the factors driving treatment recommendations. We also developed treatment outcome models for over 20 common outcome measures. These models were found to be unbiased, with reliable prediction intervals and accuracy comparable to experimental measurement error. We illustrated the application of EB-GAIT through a case study, showcasing its utility in providing treatment recommendations and outcome predictions. We then use simulations to show that combining propensity and outcome models offers the possibility to improve outcomes for treated limbs, maintain outcomes for untreated limbs, and reduce the number of surgeries performed. The EB-GAIT approach addresses the limitations of the conventional CGA interpretation method, offering a more structured and data-driven decision-making process. EB-GAIT is not intended to replace clinical judgment but to supplement it, providing clinicians with a second opinion grounded in historical data and predictive analytics. While the models perform well, their effectiveness is constrained by historical variability in treatment decisions and the inherent complexity of patient outcomes. Future efforts should focus on refining model inputs, incorporating surgical details, and pooling data from multiple centers to improve generalizability. EB-GAIT represents a significant step toward precision medicine in CGA, offering a promising tool to enhance treatment outcomes and patient care. Plain Language Summary: Clinical gait analysis (CGA) helps doctors understand walking problems in patients with movement disorders, like cerebral palsy. Traditionally, treatment decisions have relied on doctors’ experience and judgment, which can lead to inconsistent and unpredictable outcomes. To address this, we developed Evidence-Based Gait Analysis Interpretation Tools (EB-GAIT), a new method that uses advanced computer models to guide treatment decisions. What is EB-GAIT?: EB-GAIT combines two types of models: (1) Treatment Recommendation Models that predict the likelihood of specific surgeries based on historical treatment patterns at our center, and (2) Outcome Prediction Models that estimate how patients’ orthopedic deformity, walking pattern, and mobility might improve after surgery or without treatment. Using a powerful statistical tool called Bayesian Additive Regression Trees (BART), we built models for 12 common surgeries and over 20 outcome measures. These models are accurate, unbiased, and provide reliable predictions. How Does EB-GAIT Help?: EB-GAIT helps doctors understand which treatments are likely to work best for each patient. The models explain why a treatment is recommended, making it easier for doctors and patients to understand. Simulations show that combining treatment recommendations and outcome predictions can lead to better results for patients, fewer unnecessary surgeries, and more consistent care. A Case Study: We use a case study throughout the paper to demonstrate EB-GAIT using a real patient example. The models accurately predicted the outcomes of surgery, showing how EB-GAIT can support doctors in making evidence-based decisions. The Future of EB-GAIT: While EB-GAIT is a powerful tool, it doesn’t replace doctors’ expertise. Instead, it provides additional information to help doctors and patients make informed choices. In the future, we plan to improve the models by adding more details about surgeries and combining data from multiple centers. We view EB-GAIT as a big step forward in using data and technology to improve care for patients with walking problems. By making treatment decisions more precise and personalized, EB-GAIT aims to help patients move better and live fuller lives.

Background: Instrumented gait analysis (IGA) has been around for a long time but has never been shown to be useful for improving patient outcomes. In this study we demonstrate the potential utility of IGA by showing that machine learning models are better able to estimate treatment outcomes when they include both IGA and clinical (CLI) features compared to when they include CLI features alone. Design: We carried out a retrospective analysis of data from ambulatory children diagnosed with cerebral palsy who were seen at least twice at our gait analysis center. Individuals underwent a variety of treatments (including no treatment) between sequential gait analyses. We fit Bayesian Additive Regression Tree (BART) models that estimated outcomes for mean stance foot progression to demonstrate the approach. We built two models: one using CLI features only, and one using CLI and IGA features. We then compared the models’ performance in detail. We performed similar, but less detailed, analyses for a number of other outcomes. All results were based on independent test data from a 70%/30% training/testing split. Results: The IGA model was more accurate than the CLI model for mean stance-phase foot progression outcomes (RMSE_IGA = 11^∘, RMSE_CLI = 13^∘) and explained more than 1.5 × as much of the variance (R²_IGA = .45, R²_CLI = .28). The IGA model outperformed the CLI model for every level of treatment complexity, as measured by number of simultaneous surgeries. The IGA model also exhibited superior performance for estimating outcomes of mean stance-phase knee flexion, mean stance-phase ankle dorsiflexion, maximum swing-phase knee flexion, gait deviation index (GDI), and dimensionless speed. Interpretation: The results show that IGA has the potential to be useful in the treatment planning process for ambulatory children diagnosed with cerebral palsy. We propose that the results of machine learning outcome estimators—including estimates of uncertainty—become the primary IGA tool utilized in the clinical process, complementing the standard medical practice of conducting a through patient history and physical exam, eliciting patient goals, reviewing relevant imaging data, and so on.

Case: A 19-year-old female athlete experienced calf pain during sport. A complete Achilles tendon rupture was diagnosed 4 weeks after injury. Ultrasound revealed discontinuity of the Achilles tendon with 2.0 cm of diastasis, persisting in plantarflexion. Plantarflexion immobilization was initiated, and progressive dorsiflexion was used until 10 weeks from injury. At 1 year from injury, ankle magnetic resonance imaging revealed a contiguous tendon, the patient was pain-free, and had returned to high-level athletics with equivalent sport performance relative to her preoperative status. Conclusion: Certain Achilles tendon ruptures in young people may be treated nonoperatively with good clinical outcomes, even if diagnosis and immobilization are delayed and tendon diastasis persists in maximum plantarflexion.