1. Introduction
In rehabilitation, upper-extremity motor function evaluation for stroke survivors is important to plan effective rehabilitation intervention [
1,
2]. The most widely used in-person assessment in clinics is the Fugl–Meyer assessment (FMA) due to its validity and reliability [
3,
4,
5]. Despite its popularity, FMA is (1) labor-intensive and time-consuming, and (2) not sensitive enough to fine changes in motor function ability due to the coarse three point grading scheme of the FM scale [
4]. Although this grading scheme results in high inter/intra-rater reliability, it also has lower sensitivity than other clinical instruments, such as the medical research council muscle strength scale (six point scale) [
4,
5,
6]. Many clinical studies have reported this limitation, meaning that it is not possible to track fine changes of a patient’s motor function using the FM scale [
4,
7,
8,
9,
10].
Thanks to recent advances in sensor technologies, several works had reported an automated FMA system to address labor-intensiveness and time consumption issues [
11], these, however, did not attempt to propose a more sensitive FM scale using the virtue of sensor-based measurements to overcome the limitation on low sensitivity [
12,
13,
14,
15,
16,
17,
18]. It might be because most work to automate FMA showed inadequate accuracy even though the work focused on predicting the original three point FM scale. Another reason would be that the machine learning methods used in the existing works for FMA are not appropriate to handle this issue because (1) some of them (support vector machine [
17,
19] and Naive–Bayes classification [
20]) cannot be used for regression and (2) the others (extreme machine learning [
14], artificial neural network [
15], and random forest [
21,
22]) require a large amount of dimension-reduced training data for regression, which could not be collected from numerous patients in practice.
A promising solution for a more sensitive FMA is to develop a continuous scoring algorithm for sensor-based automated FMA. Meanwhile, from our first attempt to apply a sensor-enabled body tracking to the FMA automation [
17], we recently reported a sensor-based automated FMA system with a rule-based expert with binary logics originated from the linguistic grading guideline of FMA, which is different to the machine learning methods used in the other existing works [
13]. Importantly, the binary logic was verified by high grading accuracy with the original FM scale [
13] and could be applied to continuous scoring through the fuzzy logic approach, such as designing an appropriate fuzzy inference system (FIS) [
23,
24]. This FIS-based approach is promising because it (1) does not require collecting a large amount of patients’ data, (2) can consider a clinician’s ambiguous judgement mathematically [
24,
25] and (3) makes its reasoning process understandable [
24].
The goal of this study was to check the feasibility of sensor-based continuous FM scale scoring. For that, we firstly chose three representative FMA tests and developed a novel scoring algorithm for the tests based on FIS defining the fuzzy variables and rules from the FMA guideline [
23,
24,
25,
26]. Then, a sensor-based automated FMA system that can provide continuous FM scale was implemented by using the scoring algorithm and a depth sensor (Kinect V2). After investigating the achievable number of grades under the system by considering the expected error of the sensor, we showed the feasibility of the proposed scoring method through a pilot trial with nine stroke patients.
4. Discussion
In this study, we used FM7 along with FM3 to evaluate the proposed FMCA. The clinician reported that the rating of FM7 was not difficult because FM7 is a simple scale expansion of FM3. The extended scales in FM7 (0+, 1−, 1+, and 2−) appeared in 51.9% of the total FMA tests (14 out of 27). This means that there is a clear demand in clinic for evaluating motor function by using a more sensitive FM scale than the existing FM3. It should be noted that FM7 could not currently be regarded as a validated clinical tool.
The T3 FMA test resulted in lower correlation (
r = 0.903) than T1 and T2, because of a disagreement in a trial between FM
3A and FM
3 highlighted in
Table 8. The correlation becomes much higher (
r = 0.984) when this trial is excluded. Since the FM
7 of the trial were ‘2’, FM
CA had the greatest deviation for trials that belong to score ‘2’ (
Figure 4). We believe that the lower performance in T3 was caused by inaccurate tracking of the motion sensor used (Kinect V2). For T3, we extracted two
FVb features when the subject moved the hand to the knee. Here, one of the features, shoulder inward rotation ROM, could not be precisely extracted because when the subject’s distal segment of the upper limb was moving along the proximal direction, the subject’s loose patient uniform, made the measurement of the angle unreliable (about a 16 degree error) [
13]. If the proposed system was applied to 26 FMA tests, we expect that 22 of 26 tests would be free from the sensor inaccuracy problem above based on the characteristics of the inaccurate tracking investigated in [
13], except the following tests: shoulder adduction/inward rotation during hand to knee (T3), shoulder external rotation during hand to ear, forearm supination during hand to ear, and forearm pronation/supination with elbow 0°.
As mentioned, this paper proposed a novel continuous FM
CA scoring algorithm based on the fuzzy logic derived from our previous rule-based expert (binary logics). One can expect that several existing studies on automated FMA could be extended for the continuous FM scale. For instance, a linearized model that is obtained from the correlation analysis between the extracted feature (i.e., range of motion) and original FM scale rated by clinician could enable the scoring of the continuous scale [
16]. Those approaches, however, would suffer from inaccuracies due to the complexity of FMA (i.e., Pearson’s correlation coefficient
r = 0.03 in some tests [
16]), as follows. Based on the Bobath concept [
3,
4], the instructions of FMA usually ask the patient to perform a certain joint motion while constraining the other joint motions for evaluating the selective/voluntary motor performance. Hence, the FM scale is rated by clinician’s comprehensive inference based on multiple features with different types:
FVa,
FVb, and
FM, and thus it makes the dimension reduction used in those approaches (i.e., using principal component analysis [
16]) difficult. Note that this statement is supported by the complex binary logic for automating some FMA tests that were shown in our previous work [
13].
The aim of the proposed sensor-based continuous-scaled FMA system is to automate the evaluation of motor function more objectively and sensitively. From a clinical point of view, along with its convenience and time efficiency, the proposed system has the potential to improve the limited sensitivity of the conventional FM scale, which would be a novel instrument for better practice of rehabilitation. Moreover, the proposed system can contribute to effective robot-aided rehabilitation therapy due to its better sensitivity. For instance, thanks to FMCA, the intensity and difficulty of the robotic therapy can be precisely chosen, and the fine monitoring of the motor function after the therapy could be used to accurately investigate its therapeutic effect. In addition, the proposed system is promising to be utilized as a key measure for achieving precise big data for upper-extremity motor function.
This study could still be improved. We only implemented three FMA tests for the proposed continuous FM scale so as to investigate its feasibility. Since the rule-based binary logic, the basis of FIS, for most FMA tests was already found in our previous work [
13], it is promising that the unimplemented tests could be covered in a similar manner in the near future. As for the sensor system, the performance could be improved when we use a state-of-art depth sensor, such as RealSense (Intel, Santa Clara, CA, USA) or Leap motion controller (Leap Motion Inc., San Francisco, CA, USA) [
36], both of which have better resolution than Kinect only. Moreover, the reliability (consistency) test of the proposed FMCA with repeated trials and various environment would be needed to confirm the feasibility of the proposed approach. In addition, the limited number of subjects in this study could be solved through an additional clinical trial with a larger population.