Sepsis is one of the critical causes of morbidity and mortality worldwide. According to the WHO report 2020, around 11 million people died due to Sepsis. In perspective, about 1 in 5 deaths globally is due to Sepsis. (https://www.who.int/news/item/08-09-2020-who-calls-for-global-action-on-sepsis—cause-of-1-in-5-deaths-worldwide)2. Sepsis impacts almost 50 million people worldwide, every year.
Despite all developments in medical science, timely detection of Sepsis and Sepsis Management is one of the critical unsolved problems both in the developed and underdeveloped world. Maternal and child mortality-related Sepsis continues to be a reason for concern in India and other developing countries.
Early detection of Sepsis can reduce sepsis-related mortality substantially. Therefore, it is critical to identify and initiate the treatment early. Early detection of Sepsis requires continuous monitoring of physiological parameters such as blood pressure, heart rate, and many more. However, such measurements require an invasive process that has many complexities.
In this study, the researchers have attempted to develop a Machine Learning-based model that could facilitate early detection of Sepsis using non-invasive methods. The ML model developed uses the Photoplethysmogram (PPG/Oximeter) signal (non-invasive) to estimate respiratory rate, arterial inflow and predict arterial blood pressure waveform.
The researchers used MIMIC-III (Medical Information Mart for Intensive Care, version-3), a freely available critical care database for training the model and prediction. This database has information on 30,000 ICU patients. The Waveform Database recorded various physiological signals (waveforms) and periodically gathered vital signs (numerics) from the bedside patient monitors in ICUs (Intensive Care Units).
Each pulse was first segmented from a long signal sequence from the extracted PPG (photoplethysmogram) and estimated for different morphological markers such as systolic peak, diastolic peak, and dicrotic notch maximum slope point. The various PPG signal subsegments were clustered and correlated with varying clusters to clinical interpretations. The researchers conducted a temporal analysis for long-duration PPG signals using the summarised PPG subsegment representation.
The model results indicated an accurate ABP (Arterial Blood Pressure) prediction based on continuous non-invasive monitoring of PPG signals. The outcome has a tremendous implication for predicting Sepsis early on and reducing mortality. Pulse morphology stability is a non-invasive continuous monitoring system that can provide early intervention alarm and predict sepsis-related mortality.
The artificial intelligence models built as part of the project can process the data and provide inputs for critical decisions while monitoring the patients at the local level, thereby helping in a timely transfer of the sick patient to the advanced care facility.
References:
1: Pamidi AM, Joshi BB, Belde G, Goutha RR, 2021, ‘Applied Deep Learning on Pulse Plethysmography data to reduce sepsis-related Maternal and Child Mortality.’, 2021.