Predicting Hypotension by Learning from Multivariate Mixed Responses

Blood pressure is the main determinant of blood flow to organs. Hypotension is defined as a systolic blood pressure less than 90 mmHg or a diastolic blood pressure less than 50 mmHg. The severity and duration of hypotension is associated with low blood flow to organs, which often result in organ dam...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Lecture notes in engineering and computer science Ročník 2245; s. 1
Hlavní autori: Ritter, Jodie, Chen, Xiaoyu, Bai, Lihui, Huang, Jiapeng
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Hong Kong International Association of Engineers 05.07.2023
Predmet:
Accuracy
Blood flow
Blood pressure
Classification
Clustering
Hypotension
Machine learning
Missing data
Multivariate analysis
Neural networks
Organs
Probabilistic models
Vital signs
ISSN:2078-0958, 2078-0966
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Blood pressure is the main determinant of blood flow to organs. Hypotension is defined as a systolic blood pressure less than 90 mmHg or a diastolic blood pressure less than 50 mmHg. The severity and duration of hypotension is associated with low blood flow to organs, which often result in organ damage and a high mortality rate. Predicting hypotension prior to and during surgery can reduce the incidence and duration of hypotension, thus improve patient outcomes. This paper uses preoperative bloodwork and vital signs as well as perioperative vital signs in 5-minute increments as inputs to forecast hypotension. Hypotension can be represented by multivariate mixed responses for systolic and diastolic blood pressures and hypotension classification, which follow both continuous and binary distributions, respectively. The main objective of this paper is to apply a new machine learning method known as an "Interpretable Neural Network" (INN) to this clinical predictive application by simultaneously modeling mixed hypotension responses considering experts' domain knowledge. A novel data pipeline is proposed to conduct variable selection, clustering, and missing data imputation to address the issues of missing value and heterogeneous samples that are common in medical records. The customized INN method was developed and tested with a dataset containing 588 hysterectomy surgeries. Computational results suggest that the Gaussian mixture model produced better clustering results, compared to a simple clustering based on patients' lab work-up records. The novel INN method was successfully applied to the hypotension prediction, providing a prediction with reasonable accuracy and high interpretability for the prediction. The binary response has a testing accuracy of 92~95%, while the continuous responses have a root mean square error in the range of 10~25. Finally, the mixed response model outperformed the pure classification model in predicting hypotension by exploiting the hidden relationship between hypotension and the actual measures of diastolic and systolic blood pressures.
Bibliografia:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2078-0958
2078-0966