PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu 1, K. Shibai 1, Y. Hayashi 2, and K.

Transcription

1 PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu, K. Shibai, Y. Hayashi 2, and K. Tanaka 2 Graduated School of Science and Technology, Niigata University, 85 Ikarashi-2, Niigata 95-28, Japan 2 Institute of Health and Sport Sciences, University of Tsukuba, -- Tennodai, Tsukuba , Japan Abstract-The Internet technology was introduced to provide on-demand support for fitting the cycle ergometer workload personally depending on each individual s physical work capacity. It was useful to combine the ratings of perceived exertion (RPE) with objective physiological indices during the exercise. We estimated the RPE from objective indices (i.e., muscular fatigue-related indices and the heart rate), using a feed-forward type artificial neural network. Interpreting the estimation errors was useful to design an appropriate workload for an individual elderly person. Keywords - cycle ergometer, heart rate, muscle activity, ratings of perceived exertion, artificial neural network PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

2 Social Background - progression of aged society - high motivation for health Health Promotion by Exercise Wellness Purpose However physical work capacity - risk for overuse - degeneration of functions - large individual differences IT system to support Exercise for Wellness Web-based Healthcare System PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

3 Internet-based Cycle Ergometer System for Serving Personally Fitted Workload Exercise at any time and at any location Browsing by Java made programs Health Care Center Sports Gymnasium PC for maintenance & service HR sensor control info. analyzed results Monitoringfuzzy parameters If necessary Internet measured data support center Network server PC for maintenance & service Continuously Supporting at any time and at any location Internet Ref. electrode Zoom Up Home Active 4-bar electrode measured data Communication port paddle sensor Wellness Bike amp. network measurement & control unit PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

4 Workload Control for Cycle Ergometer Control by Objective Indices Heart Rate (HR) SEMG, HR Surface EMG(SEMG) - amplitude info - frequency-related info. Fuzzy control -Fuzzy Rules -Membership Functions Workload control workload WL Estimation of Total fatigue index δ Recursive Workload Control n+ = WL n δ WL WL n : workload at n-th frame WL: incremental step of workload δ: fatigue index AT controlled workload PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22 time

5 Surface Myoelectric Signals Amplitude indx m+ N ARV ( m) = emg( k) N Frequency index MPF( m) = Muscular fatigue index Correlation coefficient between ARV and MPF γ ARV-MPF f H f= f f L Objective Indices k= m f H f= f L P( m, f) Pmf (, ) m : frame number N : number of samples in a frame emg : samples of ME signals P(m, f) : power spectrum of ME signal at m-th frame f H : high-frequency limit for estimation f L : low-frequency limit for estimation Heart Rate HR frame = 5 sec PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

6 Segmentation in HR-γ ARV-MPF Scatter Graph γ ARV-MPF light seg- early seg-2 middle seg µ 2 -σ 2 µ 2 µ 2 +σ 2 last heavy Heart Rate µ 2 +σ 2 µ 2 µ 2 -σ 2 ) Weigh the random values for each sample, 2) Estimate the center of samples by fuzzy means, and tentatively segment the samples, 3) Change the weight based on the distances between the center and each sample, 4) Repeat steps ) - 3) until the center position reaches a steady point, and then 5) Divide the samples. membership functions µ, µ 2, µ 3 : mean at each segment σ, σ 2, σ 3 : s.d. at each segment PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

7 Changes in HR-γ ARV-MPF Scatter Graph γ ARV-MPF X XX X X X X X XX X X XX XX X XXXX X XX XX X (a) first set progressively increasing workload γ ARV-MPF early phase X XX X X XX X XX X XXX X X X XXX X XX X -.6 X seven months Type A (b) fourth set April, 2 November, 2 middle phase last phase subject TH (a 7-year-old man) PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

8 Objective / Subjective Representation of Fatigue myoelectric signals Subjective Representation tight electrocardiogram Objective Data RPE (Ratings of Perceived Exertion) by Borg (Med. Sci. Sports Exerc. 973) 7 Very, very light 8 9 Very light Fairly light 2 3 Somewhat light 4 5 Hard 6 7 Very hard 8 9 Very, very hard 2 RPE-Scale PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

9 Workload Adjustment Procedure Design of Control Parameters - Fuzzy Rules - Membership Functions PIW Progressively Increasing Workload CW Controlled Workload PIW next trial CW CW2 CW3 conventional approach PIW Key points classification segmentation Updating Control Parameters Key point Personal Fitting trial PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

10 Evaluation of Personal Fitting Procedure RPE HR Comparison γ ARV-MPF ε Estimated RPE Estimation ANN ε( n) = RPE( n) RPE ˆ ( n) Subjective indices Objective indices Estimation error WL WL Estimation of RPE by Artificial Neural Network PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

11 Estimation of RPE by ANN Selection of a suitable ANN Numbers of cells at each layer were 4 at the input, 4-2 at the middle, and at the output (the total number of ANNs checked was 25). - HR - γ ARV-MPF - WL - WL ε( n) = RPE( n) RPE ˆ ( n) minimum squared error Estimated RPE A suitable ANN for personal fitted workload control PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

12 Experimental Conditions - Subjects 8 senior subjects (63.5±7.5 years old) from September 2 to March 2 every 2 months 6 senior subjects (69.3±4. years old) from November 2 to December 2 every 2 weeks - Myoelectric Signals muscles: right leg vastus lateralis gain: 6dB frequency band width: 5.3Hz -.2kHz 4-bar active array electrode - Heart Rate LED-photo transistor - Metabolism lactate in blood every minute respiration gas exchange every minute - Subjective Index ratings of perceived exertion (RPE) PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

13 Workload Design for Personal Fitting workload phase phase 3 phase 5 phase 2 phase 4 phase 6 T T 2 T 3 temporary increasing workload How to determine: - level - timing WL min time AT WL max AT: Anaerobic Threshold Protocol for Exercise workload WL WL min max Little bit hard exercise and perception of workload changes AT time Conventional WL control = 7%WL Personal Fitting AT = 3%WL AT PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

14 Sufficient Workload Control WL [W] ε 5 2 WL 5 RPE ^ RPE time [frame] HR γγarv-mpf time[frame] time [frame] 2-Dec-23 RPE, RPE ^ ARV-MPF ARV-MPF γ % subj.a, Dec 23, : samples in the middle phase - HR around AT - No severer muscular fatigue Type-A, 73-years-old man WL AT = 97 W, WL LT =2 W Sufficient example for PF Subject KY PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

15 WL [W] ε Insufficient Levels and Timing WL 98[W] 6[W] RPE ^ RPE time [frame] HR γγarv-mpf time[frame] 2 5 ^ RPE, RPE WL [W] ARV-MPF ε WL 5 RPE ^ RPE time [frame] HR.5 γ -.5 ARV-MPF time[frame] 5 γ ARV-MPF RPE, RPE ^ time [frame] 2-Nov time [frame] 2-Dec-9 Insufficient example for PF Subject KY PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

16 HR-γ ARV-MPF Scatter Graphs in PF ARV-MPF γ : samples in the middle phase - HR around AT - No severer muscular fatigue subj. A ARV-MPF.5 ARV-MPF.5 ARV-MPF.5 γ -.5 γ -.5 γ -.5 (a) 2.% subj.a, Nov, Lighter than PIW (b) 6.6% subj.a, Dec 9, Relatively light for muscles, but hard for respiration (c) 33.2% subj.a, Dec 23, Appropriate exercise, comfortable Insufficient level Insufficient timing sufficient workload control PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

17 RPE RPE-Workload Scatter Graph at Personally Fitted Trial B B B B B B B BB B B B B B B B B BB B B B B B B B WL AT Workload [W] RPE estimation error J JJJJ JJJJJJJJJ JJJJJJJJJ JJ J J JJ J J JJ JJ JJ J J J EE EEE EEEEE J J JJ E JJ J E EJ E J J JJ J JJ JJJ J J EEEEEEEE EEE E EE E EE E EEEE E EEE EE E EEE EE E E EE E E E E E EEE E Workload [W] 9-94 frame frame WL [W] WL 5 RPE ^ RPE RPE, RPE ^ Type-A, 73-years-old man WL AT = 97 W, WL LT =2 W ε( n) = RPE( n) RPE ˆ ( n) 2, Dec. 23 Subject KY PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

18 Conclusions. The fuzzy rules and membership functions were designed based on the individual s physical work capacity. 2. Using a feed-forward type artificial neural network, we estimated the RPE (i.e., a subjective index) from the HR and muscular fatigue, and studied the balance between the objective and subjective representations of fatigue. 3. The gap was suitable to determine the appropriate levels and the timing of temporally inserting workloads in a basic workload variation. PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22

19 Conclusions (continued) Personally Fitting Procedure. Evaluation of basic data under progressively increasing workload. 2. Continuously changing the intensity and timing of temporary increasing workloads. 3. Comparison between subjective index (RPE) and objective indices by ANN. - small error between RPE and estimated RPE. - RPE greater than 3 and no progression of muscular fatigue. Wellness Bike for Home-Based Exercise PERSONAL FITTING PROCEDURE FOR CYCLE ERGOMETER WORKLOAD CONTROL BY ARTIFICIAL NEURAL NETWORKS T. Kiryu et al., IEEE/EMBS October 24, 22