Using explainable machine learning to identify patients at risk of reattendance at discharge from emergency departments

Short-term reattendances to emergency departments are a key quality of care indicator. Identifying patients at increased risk of early reattendance could help reduce the number of missed critical illnesses and could reduce avoidable utilization of emergency departments by enabling targeted post-discharge intervention. In this manuscript, we present a retrospective, single-centre study where we created and evaluated an extreme gradient boosting decision tree model trained to identify patients at risk of reattendance within 72 h of discharge from an emergency department (University Hospitals Southampton Foundation Trust, UK). Our model was trained using 35,447 attendances by 28,945 patients and evaluated on a hold-out test set featuring 8847 attendances by 7237 patients. The set of attendances from a given patient appeared exclusively in either the training or the test set. Our model was trained using both visit level variables (e.g., vital signs, arrival mode, and chief complaint) and a set of variables available in a patients electronic patient record, such as age and any recorded medical conditions. On the hold-out test set, our highest performing model obtained an AUROC of 0.747 (95% CI 0.722–0.773) and an average precision of 0.233 (95% CI 0.194–0.277). These results demonstrate that machine-learning models can be used to classify patients, with moderate performance, into low and high-risk groups for reattendance. We explained our models predictions using SHAP values, a concept developed from coalitional game theory, capable of explaining predictions at an attendance level. We demonstrated how clustering techniques (the UMAP algorithm) can be used to investigate the different sub-groups of explanations present in our patient cohort.

Incidence and one year outcome of periprocedural myocardial infarction following cardiac surgery: are the universal definition and SCAI criteria fit for purpose?

Introduction

The diagnosis and clinical implication of periprocedural myocardial infarction (PPMI) following coronary artery bypass grafting (CABG) is contentious, especially given its importance in the interpretation of trial data. Two accepted definitions of PPMI yield discrepant results. Little is known about the association between the diagnosis of PPMI, using high sensitivity troponin (hs-cTn), and medium term mortality in patients who undergo CABG, either alone or in conjunction with another procedure. In addition, there are currently no criteria for the diagnosis of PPMI following non-CABG surgery.

Method

Consecutive patients admitted to a cardiothoracic critical care unit (CCCU) over a six month period following open cardiac surgery had hs-cTnI assay performed on admission and every day for forty-eight hours, regardless of whether there was a clinical indication. Patients were categorised as PPMI using both the Universal Definition of MI (UDMI) and Society of Cardiovascular Angiography and Interventions (SCAI) criteria. Comorbidity data, surgical details and clinical progress in CCCU were recorded. One year mortality data were obtained from NHS Digital.

Results

There were 245 CABG patients, of whom 20.4% met criteria for UDMI PPMI and 87.6% for SCAI UDMI (figure 1). The diagnosis of UDMI PPMI was independently associated with one year mortality (hazard ratio 4.175 (95% confidence interval 1.281 – 13.608)), whereas there was no association between SCAI PPMI and one year mortality (figure 2). Of the 243 patients who had non CABG cardiac surgery, 11.4% met criteria for UDMI PPMI and 85.2% for SCAI PPMI (figure1) but neither was associated with one year mortality.

Conclusions

The incidence of SCAI PPMI in a real world cohort of cardiac surgery patients is so high as to be of limited clinical value. By contrast, a diagnosis of UDMI PPMI post CABG is independently associated with one year mortality, so may have clinical utility.

Funding Acknowledgement

Type of funding sources: Private company. Main funding source(s): Beckman Coulter - supplied the assays used in the study but had no role in the study Figure 1. Frequency of PPMIFigure 2. Kaplan Meier curves

The relationship between high-sensitivity troponin taken on admission to critical care, regardless of whether there was a clinical indication for testing, and one year mortality

Introduction

High-sensitivity troponin (hs-cTn) assays now form a key component of the diagnostic pathways for patients presenting to emergency medical services with chest pain. However, hs-cTn concentrations above the manufacturer-provided upper limit of normal (ULN) are now frequently reported in patients presenting with conditions not traditionally associated with type 1 myocardial infarction (T1MI). This is particularly true of severe illness states. We investigated the possible association between hs-cTn and 1 year mortality in critical care patients.

Method

Consecutive patients admitted to two adult critical care units (general critical care unit (GCCU) and neuroscience critical care unit (NCCU)) over a six month period had hs-cTnI assay performed on admission, regardless of whether there was a clinical indication, and the results nested unless a clinical request had been made. Comorbidity data, illness severity and critical care outcome were recorded and have been previously reported. One year mortality data were obtained from NHS Digital.

Results

After excluding patients diagnosed with T1MI by the clinical team, there were 1,033 patients remaining. At one year a total of 253 (24.5%) patients had died. The Kaplan-Meier curves in figure 1 demonstrate a positive association between mortality and increasing hs-cTnI concentrations relative to the ULN. Specifically, using the log-rank test, the mortality at one year was significantly higher (p<0.001) in patients with hs-cTnI concentrations above the ULN. Furthermore, on multivariable Cox regression analysis, the log(10) hs-cTnI concentration was independently associated with the hazard of one year mortality (hazard ratio 1.587 (95% confidence interval 1.358–1.856).

Conclusions

These data suggest that admission hs-cTnI is a biomarker for one year mortality in critical care patients. Further work is now required to assess whether any medical intervention can alter this risk.

Funding Acknowledgement

Type of funding sources: Private company. Main funding source(s): Beckman Coulter provided the assays for the tests used in this study. They had no other involvement in the study

Distribution of high sensitivity troponin taken without conventional clinical indications in critical care patients and its association with mortality

Background

High sensitivity troponin (hs-cTn) concentrations above the manufacturer's upper limit of normal (ULN) are frequently seen outside the context of MI, particularly in critical care units. The current evidence regarding the prognostic value of hs-cTn in critical care settings is discrepant.

Purpose

To describe the distribution of hs-cTn in a consecutive cohort of patients in critical care units, regardless of whether there is a conventional clinical indication, and the association of this distribution with clinical outcomes.

Methods

Consecutive patients admitted to three adult critical care units (cardiothoracic (CCU), general (GCU), neuroscience (NCU)) over a six month period had hs-cTnI tests performed serially throughout the admission, regardless of whether the supervising team felt there was a clinical indication. The results were nested and not revealed to patients or clinicians unless they were requested as part of routine care. The hs-cTnI results were correlated with parameters of clinical outcome.

Results

After excluding those diagnosed with a type 1 MI, there were 1,563 patients remaining in the study cohort (CCU 530, GCU 750, NCU 283). The median hs-cTnI was 77ng/L (IQR 11–1932ng/L, with 1081 (69.2%) patients above the manufacturer-provided ULN. Overall there was a bimodal distribution; GCU and NCU were positively skewed and CCU negatively skewed. Hs-cTnI concentrations above the ULN were associated with age, comorbidity, illness severity and need for organ support (table 1). The degree by which the hs-cTnI concentration was above the ULN remained an independent predictor of critical care mortality (figure 1) in NCU and GCU.

Conclusion

Hs-cTnI elevation taken outside the context of conventional clinical indications is common in the critically ill and is associated with age, comorbidity and illness severity. Admission hs-cTnI is an independent predictor of mortality and provides additional discriminative ability to the APACHE II score alone. This assay may represent a novel prognostic biomarker on admission in non-CCU critical care settings.

Mortality relative to ULN

Funding Acknowledgement

Type of funding source: Private company. Main funding source(s): Beckman Coulter

Inflammatory phenotyping predicts clinical outcome in COVID-19

Background

The COVID-19 pandemic has led to more than 760,000 deaths worldwide (correct as of 16th August 2020). Studies suggest a hyperinflammatory response is a major cause of disease severity and death. Identitfying COVID-19 patients with hyperinflammation may identify subgroups who could benefit from targeted immunomodulatory treatments. Analysis of cytokine levels at the point of diagnosis of SARS-CoV-2 infection can identify patients at risk of deterioration.

Methods

We used a multiplex cytokine assay to measure serum IL-6, IL-8, TNF, IL-1β, GM-CSF, IL-10, IL-33 and IFN-γ in 100 hospitalised patients with confirmed COVID-19 at admission to University Hospital Southampton (UK). Demographic, clinical and outcome data were collected for analysis.

Results

Age > 70 years was the strongest predictor of death (OR 28, 95% CI 5.94, 139.45). IL-6, IL-8, TNF, IL-1β and IL-33 were significantly associated with adverse outcome. Clinical parameters were predictive of poor outcome (AUROC 0.71), addition of a combined cytokine panel significantly improved the predictability (AUROC 0.85). In those ≤70 years, IL-33 and TNF were predictive of poor outcome (AUROC 0.83 and 0.84), addition of a combined cytokine panel demonstrated greater predictability of poor outcome than clinical parameters alone (AUROC 0.92 vs 0.77).

Conclusions

A combined cytokine panel improves the accuracy of the predictive value for adverse outcome beyond standard clinical data alone. Identification of specific cytokines may help to stratify patients towards trials of specific immunomodulatory treatments to improve outcomes in COVID-19.