327 - Noninvasive cuffless blood pressure estimation using pulse transit time-ECG monitor with photoplethysmography in pediatric patients.

Poster Number: 327
Publication Number: 327.214

Fan-Yu Yen, National Cheng Kung University, Tainan, Tainan, Taiwan (Republic of China); Rocco Landi, Boston Children's Hospital / Harvard Medical School, Boston, MA, United States; Patricia Ellen Grant, Boston Children's Hospital, Boston, MA, United States; Pei-Yi Lin, Boston Children's Hospital, Boston, MA, United States; Jason Sutin, Boston Children's Hospital, Boson, MA, United States

Background: Children who are in the intensive care unit and pre, intra, or post surgeries require continuous arterial blood pressure (ABP) monitoring. However, continuous ABP monitoring typically requires an invasive arterial catheter which imposes additional risks. Thus, there is a need to improve the non-invasive ABP monitoring methods.

Objective: We retrospectively reviewed patient monitor data to investigate using pulse transit time (PTT) to monitor ABP non-invasively. PTT is the arrival time difference between a blood pulse wave reaching two different locations in the body and is inversely proportional to blood pressure. In theory, PTT can be a non-invasive substitute for continuous ABP monitoring that could be cheaply calculated from conventional electrocardiogram (ECG) and plethysmography signals from standard patient monitors available anywhere in the world. In this study, we developed new algorithms to extract PTT features from patient monitor data and compare the estimated ABP with gold standard clinical invasive blood pressure.

Design/Methods: We collected 3342 patients’ patient monitor data with invasive arterial blood pressure monitoring since October 2020. The demographic data of the patients were: mean age ± standard deviation (SD) of 8.50 ± 9.68 years, weight of 28.84 ± 26.51 kg, and recording duration of 4.58 ± 3.40 hrs. Among all the patients, 1929 patients were diagnosed with cardiac disease. We extracted 40 features including PTT, heart rate, and various plethysmography shapes features. Four machine learning models were used to estimate ABP, including linear regression (LR), random forest (RF), feedforward neural network, and recurrent neural network (RNN). 

Results: As preliminary results, we chose 118023 heartbeats which were divided into training sets and testing sets in 8:2 ratio. The mean ± SD between LR estimated value and recorded ABP data is -2.47 ± 5.70 mmHg. The result was 1.61 ± 3.78 mmHg for RF, -0.04 ± 6.74 mmHg for feedforward neural network, and 3.09 ± 4.41 mmHg for RNN. Conclusion(s): In this study, our preliminary result established that various features from ECG and plethysmography were able to estimate ABP while the complete data analysis is still in process. Success would enable development of new metrics to extend monitoring of hemodynamic instability into patients whose condition does not justify invasive monitoring but still may be at risk of compromise.