An Improved Troika Framework for Heart Rate Monitoring Using Wrist-Type Ppg Signal＊s

XU Yi-huang

(Guiyang Preschool Education College，Guiyang 550004，China)

Photoplethysmographic (PPG)signals，recorded in a wear-able device using a pulse oximeter that illuminates a wearer’s skin using a light-emitting diode(LED)and measures intensity changes in the light reflected from skin，can be used as in-put for monitoring heart rate (HR)of the wearer since the periodicity of the PPG signal corresponds to the cardiac rhythm of heart beating.

HR monitoring is very important for exercisers to control their training load. In spite as a key feature in wearable devices，HR monitoring using PPG signals is very challenging since the motion artifacts (MA)will contaminate the PPG signals especially when the wearers are performing intensive exercise.

In the literature，a variety of signal processing techniques have been proposed to remove MA. One of the most appealing methods is the TROIKA framework proposed in［1］，which consists of signal decomposition for denoising，sparse signal reconstruction (SSR)for high-resolution spectrum estimation，and spectrum peak tracking with verification. Although good performance was reported，there is still much space for improvement. In this competition，we developed a HR monitoring system following the TROIKA framework but with several improvements that are summarized as bellow:

· To remove MA prior to the peak tracking，we further exploit the acceleration information to eliminate the MA in the spectrum.

· During peak tracking，we propose several strategies including initialization correction，tracing back and exploiting priori knowledge in the verification step，to enhance the performance of the spectrum peak tracking.

The rest of the report is organized as follows.The details of the developed system are introduced in Section 2. Section 3 presents experiment results.Conclusion is summarized in Section 4.

1 DEVELOPED SYSTEM

The overview of the developed system is shown in Fig.2. In the preprocessing stage，the raw PPG signals are preprocessed by bandpass filtering to remove noise and signal decomposition to partially MA. In the spectrum correction (SC)stage，the spectrum outputted is further corrected by removing MA using the acceleration data. Finally，the corrected spectrum is fed to Spectrum Peak Tracking (SPT)stage to compute the heart rate and output the final results.

Fig.1 Flowchart of the developed system. Our focuses are on the spectrum correction stage and the Spectrum Peak Tracking stage

Because the techniques used in the preprocessing(including band pass filtering and signal decomposition)are the same as in the TROIKA framework. In this section，we only present the details of the SC and SPT stages.

1.1 Spectrum Correction

In this section，we will further correct the spectrum by removing the MA. According to our observations，the spectrum of MA contains two peaks when the wearers are running. One contains the up-anddown movement of the wearers’center of gravity;the other represents the swing of arms. Therefore，we select two peaks of the MA as the noise sources and remove them from the spectrum outputted by the SSR stage. On the other hand，we use a signal quality measurement Q to select the distinctive peak aside from the random movement background，which can be computed as

where Apeak1and Apeak2are two most highest peaks in MA spectrum，{Ai}i=1:Nare N remaining highest peaks. We set N = 10 in our experiments.Inspired by the TROIKA framework in［1］，harmonic is also used to identify the periodical running of MA. In our implementation，after selecting the highest two peaks which Q is bigger than the threshold QT，if any two of the highest peak has a harmonic peak，then both of them will be regarded as the removable noise peaks in the SSR’s output. Now，we can use the two peaks to find the corresponding to the noise frequency of SSR’s output as FMA1and fMA2with index NMA1and NMA2. The removable noise range will be Rma1=［NMA1-δma，NMA1+δma］and Rma2= ［NMA2-δma，NMA2+δma］. To avoid mis-operation of heart rate，we also set a heart rate range RHR= ［Nprev-δHR，Nprev+δHR］where Nprevis index the previous heart rate in SSR’s output. If Rma1or Rma2has inter-section with RHR，we do not remove that peak.

1.2 Spectrum Peak Tracking

Our Spectrum peak tracking procedure basically follows the idea in［1］. But we made the following improvements:

First，to avoid the unreliable initialization value，we use the second or the third time window to initialize when the first window fails to return a correct value. According to the observation，the peak amplitudes are alike，corrupted by the noise，which should be considered as an unreliable case. Thus，firstly，determine whether the first window is reliable by choosing we find the highest peak in the spectrum and also the 10 highest peaks in the spectrum. If the any of 10 ratios of the highest and the other 9 high peaks excesses 2，then we define the current time window is unreliable as well as set the initIsReliable to“false”.When processing the second window，the initialization on the second window will be executed and replace the previous initialization value on first time window.After replace the initialization value，and use the value as Nprevto continue the rest selection and verifications in the following.

In addition，we proposed a tracing back algorithm，which is applied to trace the correct peak when it appearing out-side the range R0. When the peaks in R0are so low and flat that the highest peak in the whole heart-rate possible spectrum is 10 times bigger than it，we recalculate the R0 based on the highest peak by redefining the Nprevwith this maximum peak.Then，do the rest selections and verifications as normal cases. The detail steps are showed below:First，select the highest peak in the possible range，above which the heart rate should not appear. Then，three conditions are tested，(1)the distance of the original Nprevand the maximum peak are not too far，(2)peaks selected in R0is more than 10 times lower than the maximum peak，(3)if MA has not been removed. After all the conditions are satisfied，assign this highest peak to Nprevand redefine the range of R0and R1using this new Nprev.

Fig.2 is an example showing that the proposed tracing back algorithm significantly improves the detection results.

Fig.2 The detection results on 15-th time window of the Person 10’s training data

Finally，we add a priori knowledge in the verification step to enhance the tracking. As we know，when the wearer starts running，the wearer’s heart rate will increase and when the running speed becomes steady，his heart rate will keep. Thus we add the running status as a constrain in the heart rate estimation.In addition，we found that the harmonic information is not strong enough to utilize. Thus we enlarge the harmonic match statement in［1］’s case 1 and only search the peak near the previous heart rate in case 2.

Tab.1 The average absolute error on all 12 subjects’training data

2 EXPERIMENTS

In this section，we will report experimental results of using our developed system on both the training data (with ground truth)and testing data (without ground truth).

2.1 Results on training data

In Fig. 3，we show the detection results outputted by our sys-tem as well as the ground truth on 12 subjects’training data. As we can see that the results obtained by our method are very close to the ground truth in Fig. 4(c)，Fig. 4(d)，Fig. 4(e)，Fig. 3(g)，Fig.4(f)，Fig.3(i)，Fig.3(k). On the remaining figures，our results have some errors either on the beginning stage due to wrong initialization(Fig.4(b)，Fig. 3(d)and Fig. 3(l))or on the middle stage due to peak tracking failure caused by MA (Fig.4(a)and Fig.3(j)).

The average absolute errors on all 12 subjects’recordings are shown in Table 1. Please refer to［1］for more details about how to compute the error values.Our developed sys-tem gives 2.24 errors (type I)on average on 12 training samples，which is slightly better than the TROIKA method (2.34 reported in［1］).

2.2 Results on testing data

We also run our system on 6 subjects’testing data and show the detection results in Fig. 4. Noted that we cannot conclude whether our system works well because no ground truth is available.

3 CONCLUSION

In this contest，we developed a heart rate monitoring system，which is based on the existing TROIKA framework but with several important improvements，including 1)a MA removal approach using acceleration data，2)a tracing back algorithm that is used to trace the correct peak when it appearing outside the range，and 3)an initialization correction approach for peak tracking. Our developed system achieved better performance than the original TROIKA method on 12 subjects’training data.

Although our developed system has improved abilities of handing MA problems，experimental results also indicate our current system still has estimation errors due to wrong initialization and peak tracking failure，which are caused by severe MA. In the future work，we will continue to investigate this problem and figure out better MA removal methods.

［1］Z. Zhang，Z. Pi，and B. Liu，“TROIKA:a general frame-work for heart rate monitoring using wrist-type photoplethysmographic(PPG)signals during intensive physical exercise［J］.IEEE Transactions on Biomedical Engineering，2015，66(2):522-531，1，2，3.