Identifying Offline Muscle Strength Profiles Sufficient for Short-Duration Fes-Lce Exercise: A Pac Learning Model Approach

Abstract

Functional electrical stimulation-induced leg cycle ergometry (FES-LCE) provides therapeutic exercise for persons with spinal cord injury (SCI). However, there exists no systematic approach to predict whether an individual has sufficient thigh muscle strength necessary for FES-LCE exercise. To develop and test a Probably Approximately Correct (PAC) learning model as a predictor of thigh muscle strengths sufficient for short-duration FES-LCE exercise and compare the model's performance with other well-known statistical methods. Six healthy male individuals with SCI, having age (32.0 +/- 12.5 years), height (1.8 +/- 0.04 m), and weight (79.12 +/- 10.76 kg), participated in static and dynamic experiments. During static experiments, absolute crank torque measurements were used to estimate thigh muscle strengths in response to maximum FES intensities of 70 mA, 105 mA, and 140 mA at fixed crank positions on an FES-LCE. During dynamic experiments, changes in power output measurements were used to classify rider performance as 'Fatigue' or 'No Fatigue' during short-duration FES-LCE at maximum stimulation intensities of 70 mA, 105 mA, and 140 mA and flywheel resistance levels of 0/8th, 1/8th, and 2/8th kilopounds. A Probably Approximately Correct (PAC) learning model was developed to classify static offline muscle strength observations with online rider performances. PAC's discriminatory power was compared with logistic regression (LR), Fisher's linear discriminant analysis (LDA), and an artificial neural network (ANN) model. PAC and ANN learning models correctly identified 100% of the training examples. PAC's average performance on the validation set was 93.1%. The ANN and LR performed comparable with 92.8% and 93.1% accuracy, respectively. The LDA method faired well on the validation set at 89.9%. PAC performed well in identifying muscle strengths associated with the online performance criterion. Although PAC did not perform best during cross-validation, this model has many advantages over the other methods. PAC can adapt to changes in classification schemes and is more amenable to theoretical analyses than the other methods. PAC learning has an intuitive design and may be a practical choice for classifying muscle strength profiles with well-defined performance criteria.