Predictions are absolutely fucking useless! Give us protocols that deliver recovery and then we can discuss if you can stay employed!
A multi-modal machine learning approach to predict fugl-meyer scores and motor recovery potential in stroke rehabilitation: Toward precision-based therapies
Introduction
Stroke is a leading cause of disability, affecting approximately 100 million worldwide [1]. Most survivors suffer motor deficits (∼70 %) and require rehabilitation [2]. Effective rehabilitation, both near and long-term, should be guided by the patient’s neurological examination (e.g., motor status) and potential for recovery, which is usually informed by clinicians’ expert opinions. However, these impressions are influenced by personal experiences and cognitive biases, which explains the failure to accurately prognosticate outcomes even within the stroke specialist group [3]. In addition, current tools to predict the extent of a patient’s motor recovery or even accurately measure their current abilities remain limited. Survivors undergo clinical evaluations with a narrow focus and thus do not receive comprehensive assessments before discharge. Tele-stroke sessions may partially compensate for such issues, yet they are limited in scope of care and are neither homogenous nor yet fully developed for functional assessments [4]. This is critical because “all patients should undergo a formal assessment of the patient’s rehabilitation needs before discharge” [5]. With an aging population, a significant shortfall in US neurologists projected over the next 10 years [6], and a lack of neurologists outside of metropolitan areas [7], the challenge of providing timely and accurate assessments is growing
Traditional clinical assessments such as the National Institute of Health Stroke Scale (NIHSS – list of acronyms is provided in Supplementary Material S1) and the Fugl-Meyer (FM) scale are widely used [8,9], and may be reduced to even coarser prognostic scales (e.g., Orpington Prognostic Scale (OPS)) [10]. These scales have notable limitations. They are highly dependent on the care provider/point of care, leading to intra- and inter-observer variability and influence of assessor experience. They often omit fractionated and complex distal movements, unequally weight upper versus lower limbs [[11], [12], [13]], and fail to address real-world variability in patient conditions and environments, where motor function may differ significantly from what is observed during controlled assessments. The FM scale, while useful for comprehensively assessing post-stroke motor impairment, monitoring rehabilitation progress, and planning treatment, is subject to observer bias and inconsistency [9] and exhibits a ceiling effect in patients with mild motor impairment. Moreover, it does not account for the full spectrum of motor recovery, particularly in cases where finer motor skills and complex motor tasks are important.
The NIHSS begets inherent limitations as well [8]. For example, it sacrifices accuracy for reproducibility, under-represents certain stroke syndromes (i.e., posterior circulation, right-sided, infratentorial), lacks inter-rater reliability in some items, exhibits undesired redundancy in specific clinical contexts, presents cultural barriers, has a ceiling effect, and most importantly, like the FM, measures impairment and not disability. In summary, there is no single stroke scale that effectively captures all its complex effects [8]. Moreover, while nearly all acute stroke evaluations incorporate the NIHSS, rehabilitation trials typically use FM scores. This disconnect between acute care evaluations and rehabilitation metrics underscores the need for tools that holistically evaluate motor recovery, capturing both detailed motor performance and functional adaptability in diverse patient environments.
To address the aforementioned limitations the American Heart Association (AHA) Stroke Rehabilitation guidelines [14] specifically call for the development of “computer-adapted assessments for personalized and tailored interventions”, “newer technologies such as… body-worn sensors”, and “better predictor models to identify responders and non-responders”. However, user-friendly, clinician-tailored, patient-centered technology for a holistic real-world assessment and classification of motor symptoms and disability, as well as recovery prediction is lacking [[15], [16], [17]]. As elaborated on in Section 1.2, current solutions suffer from technical limitations and are often confined to controlled environments, as part of close-set data collection studies, which restricts their clinical applicability in real-world settings. Many of these tools rely on single-sensor modalities or fixed-point data collection methods, lacking the flexibility needed to handle the variability and noise commonly encountered in hospital settings. Thus, these systems fail to offer holistic, adaptable solutions capable of continuous monitoring, real-time assessments, and/or predictive modeling across diverse patient populations with varying degrees of stroke severity, comorbidities, recovery trajectories, and/or rehabilitation settings.
Neuroimaging, wearable and portable sensors for measuring movement kinematics and kinetics, and integrated technology approaches are increasingly used in stroke clinical assessment [18], clinical scales prediction (e.g., FM), or motor recovery prediction [15,19]. While the prognostic value of neuroimaging markers (e.g., Diffusion Tensor Imaging (DTI) measures of white matter integrity) have been highlighted, current markers cannot fully assess the functional status of preserved tissue (and thus functional outcomes) [20]. Furthermore, methods such as [15,21] that combine additional functional assessments with imaging are not widely adopted potentially given limitations such as risks and safety concerns [22]. Similarly, sensor-based kinematic and kinetic measurements are being more commonly used to quantify motor abilities (e.g., [18,[23], [24], [25], [26]]). However, many sensor technologies, such as those embedded in rehabilitation robots or complex motion capture systems, are not easily deployed in acute bedside settings due to size, cost, or complexity [27]. Wearable and portable systems have shown promise [18]. Unfortunately, though, current systems largely rely on a single sensor modality or focus on specific joints and/or symptoms, limiting their ability to capture the full complexity of motor behavior and disease state [18].
The IMAS used in this study (see detailed description in Section 2.2) integrates multimodal sensors for recording of movement kinematics and kinetics, and specifically inertial sensors, a portable camera-based system, and a portable force plate. In addition to data from these sensors, IMAS is designed to incorporate clinical information such as patient demographics, time from stroke onset, clinical assessment data, and where available, neuroimaging data (e.g., infarct volume) to derive a comprehensive assessment of patient’s motor status and potential for motor recovery.
The sensor types integrated into the IMAS have all been used separately for motor assessments in stroke, but all have specific limitations when used separately [28]. For example, inertial sensors provide only acceleration data along specific axes, with velocity and displacement estimates hindered by drift errors and offset fluctuations [29]; force plates are unable to directly measure joint position or coordination patterns [30]; and cameras are constrained by line-of-sight occlusion, sensitivity to reflections and background noise [31]. This limits its standalone utility for comprehensive motor assessment. Ultimately, systems based on a single sensor/modality have inherent limitations.
These limitations can be addressed by capturing and integrating data from multiple sensor types. A holistic view of motor function may be rendered by recordings of joint positions, movements, and coordination across the motor system. In addition, this raw data from different modalities must be synchronized, aligned, fused into a cohesive representation, and properly analyzed to prevent data fragmentation and subsequent limitations in the evaluation of a patient’s motor status and prediction of recovery potential. IMAS addresses this challenge by applying MML techniques to maximize the utility of these data types.
First, the system employs feature-extraction methods to derive metrics descriptive of upper limb movements (e.g., acceleration-based metrics from inertial sensors and position and velocity metrics from cameras). Next, IMAS uses modality fusion techniques to align and combine these metrics. Following fusion, dimensionality is assessed using Principal Component Analysis (PCA). Finally, elastic net is used both for feature selection, refining the model by eliminating less relevant features and for predicting motor impairment and motor recovery, focusing on the most predictive metrics. Herein, the system was explored using the sensor-based kinematic measures alone and coupled with the neuroimaging measures. We hypothesized that this multi-modal approach, which integrates kinematic measures from various sensor types and neuroimaging data, could predict FM scores and motor recovery, contributing to the advancement of data-driven methods for stroke rehabilitation research.
The key contributions of this work are:
- •Integrated multimodal framework: IMAS fuses clinical, wearable-sensor, and neuroimaging data via MML to predict both baseline motor impairment (FM scores) and subsequent recovery.
- •Interpretability: The system identifies specific sensor-derived metrics linked to recovery outcomes, enabling clinicians to tailor therapy to individual impairments.
- •Scalability: IMAS is designed to evolve continuously through integration with Big Data.
- •Real-world readiness: The features of portability, minimal setup time, and data fusion methods that tolerate incomplete or noisy data, make IMAS suitable for diverse clinical environments.
No comments:
Post a Comment