Skip to content

Latest commit

 

History

History
98 lines (74 loc) · 11.3 KB

BP.md

File metadata and controls

98 lines (74 loc) · 11.3 KB

Back to https://github.com/Kwangkee/rPPG

Surveys

Papers

Machine learning and deep learning for blood pressure prediction

Data availability

The experiments involved are based on a publicly available database MIMIC III, the script to download the dataset and the related experimental code are released on the Github repository: https://github.com/v3551G/BP-prediction-survey.

Multi aspects taxonomy of BP prediction methods

(1) Taxonomy 1-how to model the question of BP prediction from the perspective of machine learning?

  • assification question
  • Regression question
  • Signal conversion
  • Sequence prediction
  • Online/Incremental learning

(2) Taxonomy 2-whether feature extraction and predictive model building are performed simultaneously?

  • Machine learning-based methods
  • Deep learning-based methods

(3) Taxonomy 3-whether the relationship among different tasks is modeled?

  • Single-task learning
  • Multi-task learning

(4) Taxonomy 4-the signal source used for building predictive model.

  • Health Behavior data based
  • Trajectory data based
  • Facial video based
  • (i) mathematical/optical methods,
  • (ii) video-based ML methods,
  • (iii) video-based DL methods,

image

5.1 Critical issues

On the basis of a large number of literature analysis, we condensed six issues to be discussed, which are

  • imbalanced phenomenon,
  • interpretability issue,
  • sample duration,
  • individual diference,
  • large diference between SBP and DBP prediction accuracy,
  • and handcrafted features versus machine-learned features.

7.3 A general proposal‑towards objective evaluation of model’s performance

  • Mukkamala et al. (2021) argue that the increasing number of papers on BP prediction that pass traditional evaluation criteria are methodologically inadequate and misleading, and further revealed the capabilities and limitations of these methods based on several solid experiments. It seems that passing conventional evaluation standards (such as BHS O’Brien et al. 1993, AAMI Zhang et al. 2020b, etc.) and analysis tools (such as BlandAltman plot, etc.) may not necessarily guarantee good performance.
  • Mukkamala et al.(2021)은 전통적인 평가 기준을 통과하는 BP 예측에 대한 논문이 증가하는 것은 방법론적으로 부적절하고 오해의 소지가 있다고 주장하며, 몇 가지 확실한 실험을 기반으로 이러한 방법의 능력과 한계를 추가로 드러냈다. 기존의 평가 표준(BHS O'Brien et al. 1993, AAMI Zhang et al. 2020b 등)과 분석 도구(BlandAltman plot 등)를 통과하는 것이 반드시 양호한 성능을 보장하지는 않을 것으로 보인다.

7.4 Conclusions and future work

Future work

  • (3) It is time and necessary to explore relevant evaluation standards as well as clinical approval criterion that suitable to cuffless BP estimator, especially to those estimators based on ML and DL methods. By examining the whole process of establishing BP prediction pipeline, we revealed potential factors leading to the unreliability of results related to traditional assessment criteria such as the AAMI and the BHS standards. Besides, Mukkamala et al. (2021) has revealed the potentially misleading facts of some reported conclusions by presentating the limitations of widely-used, conventional BP evaluation standards such as AAMI, etc., and related analyzing tools such as Regression plot and Bland-Altman plot.

Conclusions

  • Meanwhile, we believe that training a general BP predictor with genuine strong generalization ability is still challenging, instead of the overly optimistic conclusions claimed in some literatures. In fact, the latest evaluation of smartphone-based BP estimator in a large clinical settings indicates that no commercialization has been made yet (Dörr et al. 2021). We appeal an objective view and deeper thinking on the reported results in a more systematic way.
  • iPhone App compared with standard blood pressure measurement –The iPARR trial. https://www.sciencedirect.com/science/article/pii/S0002870320304026

Advancement in the Cuffless and Noninvasive Measurement of Blood Pressure

4.5. Rationale behind the Necessity to Continue Research on Cuffless BP Measurement

  • Research on cuffless blood pressure and measurement has recently increased rapidly [144]. At the same time, different devices claimed to be able to measure blood pressure using different types of biomedical signals are becoming available [144,145,146]. So, with the emergence of a high number of research and devices in the market, validating the techniques and devices is becoming increasingly important [74,147,148].
  • There are several standardization protocols for BP measurement, such as AAMI British Hypertension Society Protocol (BHS), and none were intended for cuffless blood pressure measurement [95,138,140,149]. So, until there is a universally acceptable validation protocol specifically for cuffless BP measurement methods and devices, it is not easy to accept the result or evaluation of techniques from any research, irrespective of how accurate or reliable the method claims to be.
  • AAMI British Highestion Society Protocol(BHS)과 같은 BP 측정을 위한 몇 가지 표준화 프로토콜이 있으며, 커프스 없는 혈압 측정을 위한 표준화 프로토콜은 없었다[95,138,140,149]. 따라서, 커프스 없는 BP 측정 방법 및 장치에 대해 특별히 보편적으로 허용 가능한 검증 프로토콜이 있을 때까지, 그 방법이 얼마나 정확하거나 신뢰할 수 있다고 주장하든 상관없이, 어떤 연구의 기술 결과 또는 평가를 받아들이기가 쉽지 않다.

8. Challenges and Future Recommendations

  • The continuing research to get cuffless and noninvasive BP measurement model are facing challenges such as data collection, the accuracy of control data, standardization of public dataset, need for validation protocol, patient-specific issues, calibration, the efficiency of the algorithm, integration with the traditional method, issues with the specific dataset, deployment issues, and collaborations, etc. Along with discussing the challenges, there will be recommendations and ways forwards for the researcher to pursue their investigation.
  • 커프리스 및 비침습적 BP 측정 모델을 얻기 위한 지속적인 연구는 데이터 수집, 제어 데이터의 정확성, 공개 데이터 세트의 표준화, 검증 프로토콜의 필요성, 환자별 문제, 교정, 알고리즘의 효율성, 기존 방법과의 통합, 특정 데이터 세트와 같은 과제에 직면해 있다.문제, 배치 문제, 협업 등의 문제가 있습니다. 과제에 대한 논의와 함께, 연구원이 조사를 계속할 수 있는 권장 사항과 방법이 있을 것이다.

8.1. Data Collection and Accuracy of Control Data 8.2. Validation Protocol
8.3. Calibration
8.4. The Model or Technique Needs to Be Interpretable
8.6. Necessity of Collaboration with a Health Professional

guidelines for a fair assessment

  • Blood pressure estimation from photoplethysmography by considering intra- and inter-subject variabilities: guidelines for a fair assessment, https://scholar.google.co.kr/scholar?hl=ko&as_sdt=0%2C5&q=Blood+pressure+estimation+from+photoplethysmography+by+considering+intra-+and+inter-subject+variabilities%3A+guidelines+for+a+fair+assessment&btnG=

  • 많은 연구/논문들이 데이터 사용/관리에 문제가 있고, 결과적으로 과장된 결과를 보고하고 있다는 얘기. How about yours?

  • Although the use of a single PPG for assessing hypertension is promising, the relationship between PPG and BP is not completely elucidated in fact. In an attempt to indirectly solve this issue, several authors, [27]–[39], have been reporting an accuracy in line with two guidelines, one by the American Association for the Advancement of Medical Instrumentation (AAMI) and the other published by the British Hypertension Society (BHS) [40], [41]. However, as observed by Schrumpf et al. (2021) [42], there is lack of information in the papers over the distribution of the dataset used by them to ensure (1) no mixing samples of the same subject and (2) equal data size per subject. Indeed, except for [43]–[45], authors which partially or completely follow this recommendation are finding errors far above the reference values [42], [46]–[50], denoting that such an estimation problem is not resolved yet. In this work, we investigate how intra- and inter-subject variabilities in BP lead to different results of machine learning algorithms. Moreover, by considering specifically the intra-subject scenario, we compare single PPG machine learning algorithms with a regression using age, sex, weight and subject index number as attributes, and obtain similar results, indicating that the algorithms might actually be learning to identify persons and not predicting BP.

  • Alongside the previous studies, this work proposes to investigate how a single PPG signal can be used to estimate BP, with a representative set of machine learning algorithms covering both feature- and signal-based methods, but with an explicitly emphasis in the effect of the data arrangement—by considering intra- and inter-subject variabilities—in two frequently used databases: Multiparameter Intelligent Monitoring in Intensive Care II and III (MIMICII, MIMIC-III). As mentioned earlier, our work also indicates that, regarding intra-subject scenario, a regression using age, sex, weight and subject index number—therefore without a PPG signal (or its attributes) as input—already reaches an excellent performance in line with AAMI and BHS standards. Finally, we discuss the reasons for it and provide guidelines on what we believe to be essential procedures to properly evaluate BP estimation from PPG only, in order to prevent overestimation of the results.

Back to the Top
Back to https://github.com/Kwangkee/rPPG