Predicting infection and sepsis; what predictors have been used to train machine learning algorithms? A systematic review

Abstract

Abstract Introduction Sepsis is a life-threatening condition that is associated with increased mortality. Artificial intelligence tools can inform clinical decision making by flagging patients who may be at risk of developing infection and subsequent sepsis and assist clinicians with their care management. Aim To identify the optimal set of predictors used to train machine learning algorithms to predict the likelihood of an infection and subsequent sepsis and inform clinical decision making. Methods This systematic review was registered in PROSPERO database (CRD42020158685). We searched 3 large databases: Medline, Cumulative Index of Nursing and Allied Health Literature, and Embase, using appropriate search terms. We included quantitative primary research studies that focused on sepsis prediction associated with bacterial infection in adult population (>18 years) in all care settings, which included data on predictors to develop machine learning algorithms. The timeframe of the search was 1st January 2000 till the 25th November 2019. Data extraction was performed using a data extraction sheet, and a narrative synthesis of eligible studies was undertaken. Narrative analysis was used to arrange the data into key areas, and compare and contrast between the content of included studies. Quality assessment was performed using Newcastle-Ottawa Quality Assessment scale, which was used to evaluate the quality of non-randomized studies. Bias was not assessed due to the non-randomised nature of the included studies. Results Fifteen articles met our inclusion criteria (Figure 1). We identified 194 predictors that were used to train machine learning algorithms to predict infection and subsequent sepsis, with 13 predictors used on average across all included studies. The most significant predictors included age, gender, smoking, alcohol intake, heart rate, blood pressure, lactate level, cardiovascular disease, endocrine disease, cancer, chronic kidney disease (eGFR<60ml/min), white blood cell count, liver dysfunction, surgical approach (open or minimally invasive), and pre-operative haematocrit < 30%. These predictors were used for the development of all the algorithms in the fifteen articles. All included studies used artificial intelligence techniques to predict the likelihood of sepsis, with average sensitivity 77.5±19.27, and average specificity 69.45±21.25. Conclusion The type of predictors used were found to influence the predictive power and predictive timeframe of the developed machine learning algorithm. Two strengths of our review were that we included studies published since the first definition of sepsis was published in 2001, and identified factors that can improve the predictive ability of algorithms. However, we note that the included studies had some limitations, with three studies not validating the models that they developed, and many tools limited by either their reduced specificity or sensitivity or both. This work has important implications for practice, as predicting the likelihood of sepsis can help inform the management of patients and concentrate finite resources to those patients who are most at risk. Producing a set of predictors can also guide future studies in developing more sensitive and specific algorithms with increased predictive time window to allow for preventive clinical measures.