Use of Artificial Intelligence in Sepsis Detection and Management

Background

In the United States, 750,000 people are diagnosed with sepsis every year (Bansal et al., 2018), accounting for 250,000 deaths (Harrison et al., 2015) and 51% of total ICU admissions (Bansal et al., 2018). Sepsis is a complex disease with an even more complex management and treatment process. Diagnosis of sepsis consists of suspected or proven infection and systemic inflammatory response syndrome (SIRS). Delay in diagnosis places a burden on the healthcare system, costing close to 15.4 billion dollars annually

(Bansal et al., 2018). Advanced detection can help mitigate these challenges by allowing for earlier diagnosis (Harrison et al., 2015). Furthermore, initiating early treatment would have a significant impact on mortality rates and alleviate the strain on healthcare resources allocated to sepsis treatment (Bansal et al., 2018).

Sepsis Prediction

Sepsis and septic shock being complex disease entities require scanning through the large scale of data to make as a diagnosis. It often gets delayed which could be detrimental to patient outcomes. Machine learning algorithms (MLAs) could greatly benefit patient outcomes by detecting potential sepsis shock (Shimabukuro et al., 2017). MLAs can synthesize information from different variables and utilize patient data to predict specific outcomes (Nemati et al., 2018). Earlier diagnosis and utilization of sepsis bundles could lead to lower mortality rates and improve patient care (Bansal et al., 2018). MLA systems can improve the timeliness of treatment intervention by alerting the treatment team before the onset of sepsis symptoms (Bansal et al., 2018).

Most electronic health record (EHR) alert tools are score-based, relying on certain criteria to be met diagnosis (Barton et al., 2019). However, MLA is customizable. Providers can use the system to monitor specific patient measures such as lactate levels and blood pressure (Harrison et al., 2015). MLAs can provide an early diagnosis with high specificity and sensitivity which could be very beneficial in the critical care setting (Shimabukuro et al., 2017).

Sepsis Detection

Early detection of sepsis is integral to effective treatment and improved patient outcomes (Harrison et al., 2015). Compliance with bundled sepsis protocols can reduce patient mortality and improve care (Giannini et al., 2019). MLAs have the potential to relieve the burden on healthcare resources via ongoing monitoring of the patients’ risk for sepsis (Nemati et al., 2018). With machine learning systems, clinicians can monitor patient outcomes without the need for constantly being at the bedside (Nemati et al., 2018). As MLAs function as an early alert system, many healthcare providers will be able to provide prompt care well in advance (Bansal et al., 2018). There is also evidence pointing to the utilization of MLA’s in the non-ICU setting that resulted in earlier and more frequent treatment interventions in septic patients (Harrison et al., 2015).

MLAs help streamlines patient care by providing notifications only when necessary. Most healthcare professionals are often burdened with more than 50 notifications a day from EHR systems and emails (Giannini et al., 2019). This could inadvertently lead to individuals avoiding or ignoring their alerts (Giannini et al., 2019). However, MLAs can also decrease alarm fatigue by providing alerts only when necessary (Nemati et al., 2018). High sensitivity and specificity can be expected from most MLA systems that monitor key vital signs related to sepsis shock (e.g., high lactate levels, blood pressure) up to 48 hours in advance, so healthcare professionals can expect that alerts will only occur in exceptional circumstances.

A current challenge exists is that MLAs that monitor for sepsis may be alerting healthcare professionals too early (Bansal et al., 2018). A response that may occur hours in advance may result in the provider choosing not to provide treatment - especially if the patient is not exhibiting physical signs of sepsis. Furthermore, a single alert may not be enough to incur a response from the healthcare team (Giannini et al., 2019). It is unclear whether an alert system or a dynamic scoring system would be more effective in notifying healthcare professionals (Giannini et al., 2019).

Sepsis Treatment

Delayed treatment of sepsis can lead to higher mortality rates in patients (Harrison et al., 2015); however, early antibiotic therapy can improve patient outcomes (Nemati et al., 2018). The development of MLAs with high sensitivity and high specificity can improve timeliness in sepsis detection (Bansal et al., 2018). A recent study suggested that after every hour a patient is not administered appropriate antibiotic therapy, their chance of survival decreases by 7% (Shimabukuro et al., 2017). Furthermore, there is literature that suggests that timeliness of antibiotic therapy is a key determiner for sepsis outcomes (Nemati et al., 2018). Thus, early intervention is integral to mortality rates in patients undergoing septic shock. MLA systems can decrease the time between the onset of sepsis and treatment by providing alerting treatment providers promptly (Harrison et al., 2015). There is also an increase in the frequency of treatment interventions in the ICU setting when utilizing MLAs (Harrison et al., 2015).

Noncompliance with sepsis bundles can lead to increases in patient mortality (Barton et al., 2019). However, a team-based approach to sepsis management, coupled with MLA systems to alert healthcare providers before the onset of septic shock, can mitigate this risk (Harrison et al., 2015). Interdisciplinary teams that provide patient care after being notified can lead to the more frequent implementation of compliance bundles (Bansal et al.,

• 2018) . The utilization of MLAs also leads to decreases in time to ICU transfer (Giannini et al., 2019). Escalating care for patients suspected of having sepsis can occur sooner, leading to earlier treatment intervention (Giannini et al.,
• 2019) . However, it is integral that healthcare providers are informed of the variables that lead to alerts; otherwise, there may be distrust in the effectiveness of the system (Giannini et al., 2019).