Data was collected as part of Krasoulis et al. (2020). For completeness we briefly described the experimental process. The myoelectric signals were collected from the forearm of 12 able-bodied and the stump of two subjects with trans-radial limb loss. See Table 1 for a description of participants with limb loss. Prior to the placement of myoelectric sensors, we cleansed participants' skin using 70% isopropyl alcohol. In the first group, we placed 16 Tringo^TM sensors (Delsys, USA) on the participants' in two rows of eight equidistant electrodes. For subjects with limb loss, we used 13 and 12 sensors based on their physiology of the stump after amputation. We used an adhesive elastic bandage to secure the locations of the electrodes throughout the sessions.

Sampling rates in data recording were originally 1,111 and 128 Hz for the myoelectric channels and the inertial measurement data, respectively. This sampling rate was the standard and fixed hardware sampling rate of Delsys Trigno system when inertial data was recorded simultaneously. We did not use the recorded inertial data in this study. In the process of data cleaning, the Hampel filter was applied to up-sampled (2 kHz) data to eliminate the power line inference. This pre-processing followed the process introduced in Atzori et al. (2014). The data was then band-pass filtered in the range 30–400 Hz using a 4th-order Butterworth filter. In this paper, we used only the myoelectric data. In a sliding window of 128 ms (256 raw signal samples) with an increment of 50 ms (100 raw signal samples) the waveform length for each channel was calculated resulting in 16 values per window (one per sensor), and 13 and 12, respectively participants with limb loss. In the discussion, we described why we chose only one feature.

Data was collected in two identical sessions, namely T1 and T2, which were conducted sequentially on the same day. During the recording session, each subject was instructed to perform the grips that were shown on a computer screen. The experiment involved five grips, namely power grip, lateral grasp, tripod grasp, index pointer, and hand open. During two blocks of data recording, each grip was repeated 10 times, with 5 s of muscle activation followed by 3 s of rest.

The first dataset, T1, was used for training and the second dataset, T2, was used for testing the decoder.

A general framework for an active learner includes two sets of data, namely an initial small set of labeled data L and a large pool of unlabeled samples U. Based on initial subset of data, L^* ⊂ L, a model f(x|L^*) is trained. The aim of an active learner is to choose an L^* such that f(x|L^*) ≈ f(x|L). Note that these two sets are not the same sets as T1 and T2.

To obtain L^*, the active learner adopts a query strategy and a current model f(x|L′), where L′ is a transitional annotated set, and selects new samples for annotation by an oracle, e.g., the prosthesis user. The oracle has the task of labeling the unlabeled data. Figure 1 presents a generic active learning framework. Starting from the initial dataset, the model is trained. Adopting on a query strategy, the active learner selects new instances x_U from the pool of unlabeled data for labeling by the oracle and moving them from U to L. The model is then updated.

There are three main strategies of sampling from the pool of unlabeled data, namely, stream-based selective sampling (Atlas et al., 1990), membership query synthesis (Angluin, 1988), and pool-based active learning (Lewis and Gale, 1994). We adopted the pool-based sampling method in this proof of concept, offline, study. We also shared our views as to how active learning can be implemented in a real-time myoelectric control setting.

Active learning is based on defining optimal criteria for deciding whether the oracle needs to label an unlabeled sample, that is a query. Using measures that determine the informativeness and/or representativeness of a sample, there are many algorithms for querying the label. Informativeness is typically measured with uncertainty sampling. In this method, the active learner finds samples in the pool about which the decoder is least certain. We used the following metrics:

Alternative to the single sample selection, a batch-mode active learning can be applied, which allows querying larger sets of samples. We used ranked batch-mode active learning (RBMAL) (Cardoso et al., 2017), which generates an optimized ranked list (Q) of unlabeled samples. Specifically, the uncertainty estimation step calculates two sets. First, uncertainty score U_score for all unlabeled samples in U, and D_estimate set with L samples and already included (if any) samples within the ranking Q. The set D_estimate presents the expected training set given all samples within the ranking Q will be annotated. Then, for instance, ranking score value for each sample within U (Equation 4) is calculated using similarity score Φ and U_score. To define Φ, the highest similarity between samples from the expected training set D_estimate and each U instance was calculated. Combining this information enables the calculation of score to determine an instance with the highest value to be added to Q (and removed from U). This process for ranking construction is repeated until U = ∅ or |Q| = batchSize.

where α=|U||U|+|L′| balances the trade-off between exploration and exploitation.

We ensured that all classes, including the rest and grasp classes, had the same number of samples, i.e., balanced classification. This was achieved by under-sampling of the rest class which appeared between each grip repetition. The initial training set L ⊂ U included 300 samples in total, that is 50 samples per class. We sought to simulate a real-life scenario and hence picked the first 50 samples of data in each class in T1. This is equivalent to ~2.6 s of myoelectric data per class, in the time domain. We treated the rest of the data in U as unlabeled samples. All training data was from the first recording session, T1.

For single instance sample selection we compared the four approaches, namely random sampling (i.e., passive learning), least confidence uncertainty, smallest margin uncertainty, and entropy reduction. We used the results of the random sampling approach as benchmark. We made 1,500 queries iteratively and at each iteration the training set was normalized to its z-scores. Then a linear discriminant analysis model was fitted on the normalized training set, L′. All resultants 1,500 models were tested with the data from the second session, T2.

The second analysis compared the performance of single instance sampling with that of two batch-mode sampling methods, namely naive and ranked. The former approach focuses on finding n best samples to query within single iteration. The latter learns to obtain an optimal set of samples to query from. The ranked method removes potential sample redundancy within a queried batch (if queried samples are too similar) by emphasizing sample diversification using the Euclidean distance measure.

To avoid any bias in training the decoder using the random batch sampling method, we imposed the constraint of sampling one sample per class within the same batch, keeping the batch size to 6. This ensured an independent and identically distributed random sampling condition, imposed on random sampling scenario, throughout all the iterations. All previously mentioned query strategies were used in this analysis too.

Metrics used for the evaluation of the proposed approaches included accuracy and F1-score (i.e., harmonic mean of precision and recall). Accuracy is the percentage of samples that have their labels correctly recognized. F1-score was calculated per class using one-vs.-all scheme and then averaged across the classes, that is micro-averaging. These measures, together, reflect the overall success of the proposed approach.

Four query strategies were compared across 1,500 single queries (equivalent to 3.2 min of data). Figure 2 presents the accuracy scores of all query strategies plotted against a number of queried samples and averaged across all participants. The classifier was initially trained with 50 samples from each class (giving 15.9 s of training data per subject). The mean accuracy and standard deviation for able-bodied participants and people with limb difference, before the first query was 63.66 ± 16.13 and 54.24 ± 2.65%, respectively.

Analysis of data from able-bodied participants showed that, active learning, irrespective of the adopted sampling method, outperforms random sampling by 4–5%. This improvement for people with limb difference was ~2%. Although entropy sampling outperformed the baseline, it required more samples to level off with remaining active learners. These improvements might seem small, but we would like to remind the reader that they were achieved with only 1,500 samples (only 3.2 min of data) and very little additional computations. For completeness we also considered the case of 200 queries only. Details on obtained results for both query numbers and all sampling strategies are presented in Table 2.

Figure 3 depicts accuracy scores obtained by all the subjects after querying the first 200 and after all 1,500 samples. For a comprehensive analysis, all combinations are reported. In the first row, three adjacent plots map out random sampling against active learning query strategies. In all cases, active learning outperformed random sampling in almost all subjects, that is markers representing accuracy results from individual participants lie above the unity line in Figure 3A. When active learning approaches were compared against each other, results were relatively similar (c.f. Table 1). Interestingly, for most subjects, margin sampling proved slightly more successful.

Figure 4 represents number of queried instances from each grasp per query strategy (Figure 4A), averaged across all able-bodied participants. In random sampling, the data is picked with discrete and uniform probability distribution across all the classes, hence when 1,500 samples is being queried, ~250 samples is acquired from each grip class. These are then averaged across all participants. However, with active learning, we can observe a class-specific sampling. When inspecting grip trends for active learning query strategies, we observed rest class being acquired the most by entropy sampling for both participant groups. Characteristic for all strategies, most samples were queried for the lateral grasp class. This is explainable from an anatomical point of view: the abductor pollicis brevis longus muscle, which is responsible for thumb movement and stabilization, is a deep muscle in the forearm. Hence the myoelectric data from this muscle is affected significantly by volume conduction and interference from other muscles. As such active learning prioritized querying this class.

Figure 4B shows the same analysis on data, averaged across the two people with limb difference. The most striking difference between data from the two groups was that active learning required significantly smaller number of samples for the rest class. Yet, the inter-subject differences were too large to warrant statistical significance.

For batch mode analysis, two scenarios were considered: naive batch sampling and ranked batch sampling. Both approaches followed similar protocol for experimental design as in single instance sampling. Active learning framework was constructed with a linear discriminator analysis decoder, 50 initial samples (acquired in the same manner as in the previous experiment) and 250 queries (with six samples per batch, corresponding to 3 ms) which is the equivalent of the number of samples obtained after 1,500 queries for single instance sampling (250 queries × 6 samples).

Observed results in Figure 5 indicate single instance sampling and batch mode sampling behave in a similar manner. Margin and uncertainty sampling perform the best. However, ranked batch mode sampling presents divergent results. In the initial stages of querying, random batch sampling outperformed active learning query strategies. Entropy and uncertainty sampling slowly converged and outperformed random batch sampling (as the number of iterations increased). Important to note, random batch sampling was constraint to querying a sample from each class (within a single batch) at each iteration to ensure no bias across classes occurs in further iterations. Based on this constrained, batch size for sampling new instances was set to six (number of classes) for all query approaches. In contrast with the previous approaches, margin ranked batch sampling had the lowest performance across all sampling methods.

Creating a human-in-the-loop environment has the potential to motivate the users to engage more with the decoder and ultimately lead to an enhanced control of the prosthesis. We presented a closed-loop setting in which an oracle and a prostheses user are the same. Figure 6 offers one perspective by which an active learning paradigm could be implemented in real-time.

One might assume that asking the user for specific labels in a real-time setting, e.g., in the stream-based active learning paradigm, can prove difficult and indeed no hardware setting is readily available to capture such user input. To minimize user input when the active learner is requesting incoming data to be labeled, the user has to specify only whether the decoder prediction was correct or not (a binary choice). Real-time prosthesis movement is performed only after unlabeled samples from movement intention pass safety check (certainty/confidence criteria) or the user has confirmed the decoder prediction. The additional R&I criteria on incorrect predictions (false positives) would work as a filter for noise/outlier detection and sample re-selection for weak supervision.

This way samples will augment additional information and can be re-integrated into the pool. Furthermore, this approach has the potential to minimize the annotation/labeling noise. It is possible that additional steps may be required to resolve issues around the user bias.

A limitation of this work is in using a simulated human annotator as an oracle for the active learning framework; meaning the entire dataset contains pairs of {x, y} data. Nonetheless, labels were introduced to the learner only when requested by query strategy.

In this study, we focused on the informativeness of the samples. Further experiments with alternative query strategies may be needed for more detailed understanding of the potential benefits of active learning for prosthetic control. Considering long-term adaptation of the prosthesis in the real-world, the exploration versus exploitation paradigm could be further investigated. The trade-off between data informativeness and representativeness could potentially enhance the decoder functionality. While initially greedy feature space exploration enables the model to rapidly boost its performance, with time defining specific grip characteristics could possibly establish its singularity aspects.

Data used during this study presented an example of a high-quality dataset with multiple subjects including people with limb differences. From each participant, data was recorded in two distinct, but consecutive, sessions in 1 day. Between the two sessions, the electrodes remained stuck to the skin with the help of adhesive tape. This arrangement minimized the between-session variability in the data, as one would expect to see with dry electrodes especially when considering the donning and doffing of a prosthesis socket. This is neither a limitation of the proposed active learning method nor a shortcoming of the paper. Nevertheless, this repeatable recording method minimized the additional benefit of using active learning.

Commercial myoelectric controllers such as Sense (IBT, USA), Myo Plus (Ottobock, Germany), and CoAPT Gen2 (CoAPT, USA) provide supervised recalibration options with varying levels of user-friendliness. CoAPT also offers Adaptive Advance™as an automatic, continuous-learning algorithm that combines new calibration data into the existing state of users' control. However, no technical information about this adaptive calibration method is available, beyond the generic text in paragraph [0011] of CoAPT pattern recognition patent (WO2020172261).