An incremental explanation of inference in Bayesian networks for increasing model trustworthiness and supporting clinical decision making.

Various AI models are increasingly being considered as part of clinical decision-support tools. However, the trustworthiness of such models is rarely considered. Clinicians are more likely to use a model if they can understand and trust its predictions. Key to this is if its underlying reasoning can be explained. A Bayesian network (BN) model has the advantage that it is not a black-box and its reasoning can be explained. In this paper, we propose an incremental explanation of inference that can be applied to 'hybrid' BNs, i.e. those that contain both discrete and continuous nodes. The key questions that we answer are: (1) which important evidence supports or contradicts the prediction, and (2) through which intermediate variables does the information flow. The explanation is illustrated using a real clinical case study. A small evaluation study is also conducted.

Managing the surge in demand for blood following mass casualty events: Early automatic restocking may preserve red cell supply.

BACKGROUND: Traumatic hemorrhage is a leading preventable cause of mortality following mass casualty events (MCEs). Improving outcomes requires adequate in-hospital provision of high-volume red blood cell (RBC) transfusions. This study investigated strategies for optimizing RBC provision to casualties in MCEs using simulation modeling. METHODS: A computerized simulation model of a UK major trauma center (TC) transfusion system was developed. The model used input data from past MCEs and civilian and military trauma registries. We simulated the effect of varying on-shelf RBC stock hold and the timing of externally restocking RBC supplies on TC treatment capacity across increasing loads of priority one (P1) and two (P2) casualties from an event. RESULTS: Thirty-five thousand simulations were performed. A casualty load of 20 P1s and P2s under standard TC RBC stock conditions left 35% (95% confidence interval, 32-38%) of P1s and 7% (4-10%) of P2s inadequately treated for hemorrhage. Additionally, exhaustion of type O emergency RBC stocks (a surrogate for reaching surge capacity) occurred in a median of 10 hours (IQR, 5 to >12 hours). Doubling casualty load increased this to 60% (57-63%) and 30% (26-34%), respectively, with capacity reached in 2 hours (1-3 hours). The model identified a minimum requirement of 12 U of on-shelf RBCs per P1/P2 casualty received to prevent surge capacity being reached. Restocking supplies in an MCE versus greater permanent on-shelf RBC stock holds was considered at increasing hourly intervals. T-test analysis showed no difference between stock hold versus supply restocking with regard to overall outcomes for MCEs up to 80 P1s and P2s in size (p < 0.05), provided the restock occurred within 6 hours. CONCLUSION: Even limited-sized MCEs threaten to overwhelm TC transfusion systems. An early-automated push approach to restocking RBCs initiated by central suppliers can produce equivocal outcomes compared with holding excess stock permanently at TCs. LEVEL OF EVIDENCE: Therapeutic/care management study, level IV.

Applicability of Glasgow Aneurysm Score and Hardman Index to elective endovascular abdominal aortic aneurysm repair.

OBJECTIVE: This retrospective study aimed to explore the role of Glasgow Aneurysm Score (GAS) and Hardman Index (HI) in predicting outcome after elective endovascular aneurysm repair (EVAR). METHODS: All 71 patients who underwent elective EVAR in a single centre over 9 years were reviewed. Clinical data were used to classify patients into the three standard GAS tertiles and to score patients according to the HI. RESULTS: Fifty-one patients scored > or = 77 according to GAS. Actual and predicted mortality in this group were 3.9% and 9.3%. Seventeen patients scored between 69 and 77 with actual and predicted mortality of 0% and 4.1%. Three patients scored less than 69 with actual and predicted mortality of 0% and 2.4%. Ten patients scored > or = 3 on the HI with actual and predicted mortality of 10% and 100%, respectively. Twenty-four patients scored 2 with actual and predicted mortality of 4.2% and 55%. Twenty-seven patients scored 1 with actual and predicted mortality of 0% and 28%, respectively. Ten patients scored 0 with actual and predicted mortality of 0% and 16%, respectively. The chi(2) test showed extremely significant p value of 0.0001 in case of HI, and p value of 0.0800 for GAS, slightly less significant, probably due to the small sample size. CONCLUSION: Contrary to their role in ruptured and open aortic aneurysm repair, GAS and HI overestimate both mortality and morbidity following EVAR and are poor predictors of outcome.

An assessment of covalent grafting of RGD peptides to the surface of a compliant poly(carbonate-urea)urethane vascular conduit versus conventional biological coatings: its role in enhancing cellular retention.

The aim of sodding prosthetic grafts with endothelial cells (EC) is to establish a functioning antithrombogenic monolayer of EC. Application of basement membrane proteins improves EC adherence on ePTFE grafts. Their addition to a biodurable compliant poly(carbonate-urea)urethane graft (CPU) was studied with respect to EC adherence. Preclot, fibronectin, gelatin, and collagen were coated onto CPU. RGD peptide, heparin, and both RGD and heparin were chemically bonded to CPU. Human umbilical vein EC (HUVEC) labeled with 111-Indium oxine were sodded (1.8 x 10(6) EC/cm(2)) onto native and the modified CPU. The grafts were washed after 90 min and EC retention determined. The experiments were repeated six times. EC retention on native CPU was 1.0 +/- 0.2 x 10(5) EC/cm(2). The application of preclot, fibronectin, gelatin, and collagen did not improve EC retention, which was 0.8 +/- 0.1, 0.4 +/- 0.1, 0.3 +/- 0.08, and 0.5 +/- 0.2 x 10(5) EC/cm(2), respectively. Bonding RGD, heparin, and both RGD and heparin significantly improved EC retention to 1.9 +/- 0.6, 1.7 +/- 0.5, and 2.6 +/- 0.6 x 10(5) EC/cm(2), respectively (p < 0.01). Bonding of RGD, heparin, and both RGD and heparin accelerates and enhances EC retention onto CPU. Simple coating of basement membrane proteins confers no advantage over native CPU.