SepINav (Sepsis ICU Navigator): A data-driven software tool for sepsis monitoring and intervention using Bayesian Online Change Point Detection

SepINav (Sepsis ICU Navigator): A data-driven software tool for sepsis monitoring and intervention using Bayesian Online Change Point Detection

Introduction

The alarmingly rising rates of sepsis and septic shock, as well as the related mortality, morbidity, and yearly treatment expenditures among ICU patients, are a growing source of worry. SepINav is a medical informatics project that enables ICU practitioners and researchers to more efficiently and interactively monitor and intervene on current sepsis patients, as well as perform retrospective studies to find rationales for various sepsis scenarios in the ICU.

Furthermore, Bayesian Online Changepoint Detection will aid practitioners in recognising structural alterations in patients’ vital sign regimes that might signal septic shock. In addition, several additional features are added to this data-driven software tool to promise efficient monitoring and intervention to address confounding medical interventions in the ICU. The first section concentrates on the aspects of the software architecture, features, and functionalities. The next section delineates illustrative examples and discusses how this application can help the clinicians. The latter part  includes the scientific and broader impact is portrayed.

Software description

Sepsis is a broad term applied to capture a process that is not entirely understood yet. The task force of 2016 convened by the Society of Critical Care Medicine and European Society of Intensive Care Medicine addressed the pragmatic compromise and emphasized on generalizability and readily measurable identifiers to define sepsis-3. Hence, both the qSOFA and SOFA reflect the current conceptualization of underlying dynamics using physiologic and biochemical tests offered in routine patient care. SepINav identifies a patient by its hospital admission id.

This application, plugged into hospital EMR, can extract patients’ information, vital signs, lab tests, interventions, and prescription records for the purpose of monitoring. Multiple interventions at the same time are represented separately to avoid any confusion. Immediately below the intervention visualization, practitioners can observe the patients’ condition in the sepsis spectrum. As some hospitalizations and practitioners have not adopted the recent theoretical advances in the sepsis diagnostic criteria yet, this application offers practitioners to observe patients’ conditions in both Sepsis-2 and Sepsis-3.

For Sepsis-3, the spectrum has three clinical states: Non-sepsis, Sepsis, and Septic Shock. On the other hand, the Sepsis-2 spectrum involves four clinical states: Non-sepsis, Sepsis, Severe Sepsis, and Septic Shock. This application strictly follows the criteria provided for Sepsis-2 and Sepsis-3 in the consensus conference in 2001 and 2016, respectively. Fig. 1 illustrates the algorithm used to detect different clinical states for Sepsis-3 (Figure 1).

This application facilitates practitioners and research by identifying points where structural changes occur in the data stream for all the vital signs and critical parameters associated with sepsis-2 and sepsis-3. To select the changepoint detection algorithm for this application, we had several criteria. First, the algorithm has to detect the changepoint in the physiological parameters online; its accuracy needs to be competitive with the available offline change point detection algorithms.

Second, the algorithm should have the provision to take in the domain knowledge information as input so that considering this domain information can reflect in leveraging with the changepoint detected. Fifth, the change point detection algorithm has to detect any kind of change in time-series data, not specific to mean only or variance only. Our comprehensive literature study finds that only Bayesian Online Change Point Detection algorithm meets all the criteria.

Here, each node includes five state values. First, the probability density function (pdf) value of the new datum; second, the new distribution values computed, factoring the new datum in underlying distribution; third, Hazard value. Fourth, growth probability of run-length represented as the multiplication of three probabilities: the probability of the continuation of the run-length till “r-1”, the probability that no Hazard happened till arriving the run length to “r”, pdf for the new datum in terms of the underlying probability distribution; and fifth, the change probability value for that particular node. If the EHR systems provide very frequent bedside vital signs data, some anomalous data may affect the result of our interest in detecting change points. In that case, we take the average of every ten-minutes’ data to eliminate any unintentional noise in terms of the change point detection.

Illustrative examples

For illustrative examples on this application’s functionalities, we use MIMIC-III (Version 1.4): an extensive, single-center, and a comprehensive database comprising information pertaining to patients admitted to the critical care units at Beth Israel Deaconess Medical Center in Boston. Figure 2 shows a segment of the interface that facilitates monitoring the patient’s condition in Sepsis-2 spectrum and practitioner’s interventions in detail throughout the entire trajectory.

In the top left corner, there is an option to select a patient with his hospital admission id and an option for navigation, such as Sepsis-2, Sepsis-3, and correlating vital signs. In the top right corner, one can observe the preliminary information about a patient collected during admission.

Figure 3 shows a segment of the interface that facilitates monitoring the patient’s condition in the Sepsis-3 spectrum (Non-sepsis, sepsis, and septic shock) and practitioner’s interventions in detail throughout the entire trajectory.

In both interfaces offering Sepsis-2 and Sepsis-3 navigation, this application provides visualizing the individual vital signs’ and critical parameters’ trajectories associated with Sepsis-2 and Sepsis-3, respectively, integrated with the intervention timeline.It will help to study the impact of an intervention on the particular physiological parameters. Besides, the contextually-tailored bayesian online change point detection algorithm will provide the points where structural changes occur in the data stream for all the vital signs and critical parameters associated with sepsis-2 and sepsis-3.

Impact and sustainability

SepINav embodies a step towards bridging the perceptual gap between explainability and computational intricacy centered around machine learning-based medical informatics solutions. It is represented by the facility to navigate both patients’ trajectory and the interventions made by the practitioners and to detect the changepoints in vital signs that may harbinger prior to septic shock onset. This disease-specific and custom-tailored tool has a considerable impact on ICU monitoring and interventions, as sepsis is shaped by both host factors and pathogen factors in a convoluted manner and resulted in aberrant host response.

This effort will help practitioners navigate how their clinical interventions reflect in patients’ responses in vital signs and sepsis spectrum. Besides Sepsis-2 and Sepsis-3, it will help practitioners to navigate patients’ situations in terms of the common vital signs available in bedside monitoring and to notice whether there is any structural change in patients’ trajectory. Apart from that, since this tool is fueled by electronic medical records (EMR) data, it opens up two scopes of functioning: first, helping practitioners in monitoring and intervening on the existing sepsis patients, and second, conducting retrospective studies on the previous patients for research purpose and seeking rationales to different sepsis scenarios confronted in the ICU.

From the sustainability aspect of the solution, changepoint detection has the potential to address challenges and pitfalls confronted while developing machine learning-based predictive models using EMRs. In a certain way, this application can inspire new and better software to address different challenges and pitfalls in patient monitoring and risk stratification.

Conclusion

It makes navigation easier since practitioners can see the impact of their efforts on patient’s sepsis symptoms.  We demonstrated SepINav, a data-driven software solution that helps ICU practitioners to navigate patients’ sepsis trajectory, taking into account both sepsis-2 and sepsis-3. The structural abnormalities in patient’s frequently alert practitioners prior to the development of septic shock. Using Bayesian Online Change Point Detection, this tool aids in the collection of these changepoints. This programme proven its use in-

  • monitoring and acting on existing sepsis patients and
  • conducting retrospective investigations for research objectives and exploring rationales for various sepsis scenarios seen in the ICU.

Source:  Nazmus Sakib, Shiyu Tian, Md Munirul Haque, Rumi Ahmed Khan, Sheikh Iqbal Ahamed, SepINav (Sepsis ICU Navigator): A data-driven software tool for sepsis monitoring and intervention using Bayesian Online Change Point Detection,SoftwareX,Volume 14, 2021, 100689, ISSN 2352-7110, https://doi.org/10.1016/j.softx.2021.100689.