Description

DELIVERABLE 1: Overview of Data Collection Plan. Subacute stroke survivors (N=400) will be recruited and randomly assigned to one of four groups, which will receive usual rehab care only, and usual care plus acupuncture, robotic, or neuromuscular electrical stimulation (NMES) training, respectively. Each subject will be evaluated before (A0), after (A1) and 6 months after the start of intervention (A2). At A0, thorough clinical, neurophysiological (EEG, EMG, TMS), MRI, and blood-based assessments will be made for deriving predictive recovery markers. At A1 and A2, only the clinical scores and EMG will be assessed for characterizing post-rehab functional gain and long-term recovery. Also, in all groups clinical scores will be recorded midway through intervention (A½) to monitor treatment progress. Treatments will last ≥1 month with ≥20 training sessions. All A0, A1, and A2 data will be analyzed offline with machine learning and other methods, and then used to train the AI model, PRAISE-HK (Precision Rehabilitation AI System for Enhancing recovery in Hong Kong and beyond).

DELIVERABLE 2: Explainable AI Models of Recovery. Overview. Our AI system receives high-dimensional data inputs from diverse modalities, each of which requires unique processing techniques tailored to the specific characteristics of the data type for analysis. Neuroimaging data, for instance, demand specialized image processing algorithms for deriving brain structural and functional measures, while time series data such as EEG and EMG must first be analyzed with signal processing tools that capture embedded patterns across varying temporal resolutions. The need for domain-specific preprocessing of diverse data types engenders a level of analytic complexity that defies standard deep learning techniques, which may not accommodate the unique challenges presented by each modality. To obtain actionable, clinically relevant insights while ensuring interpretability and integration of the multimodal data, more sophisticated algorithms beyond standard deep learning are needed.

As such, the investigators propose the construction of a sophisticated AI fusion model as the core of PRAISE-HK. Fusion model is a machine learning approach that integrates input features from multiple data modalities in a way that leverages their unique contributions to make a prediction more accurate and robust than any single modality could produce. Although the number of subjects here (N=400) is, to the best of our knowledge, the highest among all similar studies, for high-dimensional data this number is small enough to impose rather stringent constraints on the choice of analytical methods should overfitting be avoided, and model generalization ensured. With these considerations, the investigators will construct our fusion model in two phases. Phase 1 involves extraction of predictors from the individual modalities through their independent processing with domain-specific AI models. This phase is necessary since each modality can yield the most valuable information only with distinct processing pipelines. Phase 2 involves the fusion of these processed modalities. The investigators will formulate a specialized fusion model that accommodate the relatively small sample size, the predictors' high dimensionality, and potential instances of missing data. The final model will capitalize on the strengths of the domain-specific phase-1 outputs and combine their insights for improved prediction accuracy.

Phase 1: Modality-specific Feature Extractions. For the neuroimaging and multi-omics modalities, the investigators will leverage pre-trained models to map high-dimensional inputs to a reduced set of informative features or embeddings. The investigators will start by fine-tuning foundation models based on each dataset, and then develop modality-specific models using the pre-trained and fine-tuned foundation models as feature extractors. For time-series data such as EMG, EEG, and motion capture data, both standard and custom-derived procedures will be used to filter noise and identify suitable predictors. From the EMG, non-negative matrix factorization and our recently proposed rectified latent variable model will be used to extract muscle synergy indices, which, from our preliminary results (see below), contain predictive stroke recovery information. From the EEG and EMG, coherence between these two signals (cortico- muscular coherence) and that between EEG and muscle synergy activations (cortico-synergy coherence) will be computed through spectral analysis. From the motion capture data, movement parameters such as joint motion range can be extracted with standard methods. All parameters above will be used as predictors in phase 2.

Phase 2: AI Fusion Model. Once the investigators have extracted the predictors from each modality, the investigators will develop a fusion model, one for each treatment group and each clinical score, that integrates these A0 predictor inputs for a prediction of the A1 and A2 scores . Since the investigators do not have prior information on the suitability of the different methods to be used in fusion models, the investigators will evaluate the performance of various models. Artificial neural networks with meta-learning will be used for their capacity to learn from a rich set of data features. Their performance will be benchmarked against other models such as logistic regression, Bayesian methods, and random forests, methods that are advantageous for smaller datasets such as ours. For each treatment group, the fusion model will be trained on the phase-1 predictors with the labelled data from the 100 subjects as inputs to predict the score improvements at A1 and A2.

Realistically, the investigators anticipate that our final database will have occasional missing data points. The investigators will employ mutual information-based imputation techniques to estimate missing heterogenous values, thereby making full use of the available data and preventing biases in model training. Sensitivity analyses will be conducted to assess the impact of the missing data on model predictions, thereby ensuring the reliability of our findings.

Through the entire system construction process, the investigators will utilize and compare different AI algorithms for phase-1 feature extractions and phase-2 fusion. The models and approaches employed will be iteratively refined based on test performance metrics and clinical feedback. Historical data will be utilized to validate the models through retrospective assessment of the effectiveness of the recommended interventions in previous clinical scenarios.

Additional Phase 3: Building Decision-making Framework. For every subject, by comparing the phase-2 outcome predictions across the treatment-specific models, one may already decide on an ideal intervention for the subject simply by selecting the treatment whose model yields the best A2 improvement. But this selection process is not transparent, because the fusion models do not explicitly indicate how the many predictors interact to lead the models to collectively arrive at the recommended treatment. To enhance the AI system's explanability, the investigators will implement a phase 3 for constructing an explicit decision-making framework that guides intervention selection, one analogous to the clinically successful PREP decision-tree algorithm of Stinear et al. A classifier will be trained with the phase-1 predictors and the phase-2 score predictions of all subjects (N=400) serving as inputs, and the ranked intervention preferences inferred from phase-2 outcome comparison as outputs. The trained classifier can serve as a decision support tool that reveals the most parsimonious set of evaluations needed for reaching decisions, and provides transparent reasoning behind its recommendations, thereby enhancing our understanding of why particular subjects are assigned to each treatment. To maximize interpretability, the investigators will implement decision-tree algorithms such as the random forests, which are capable of discerning complex, nonlinear relationships among the variables.