Where multiple samples were available in a single patient, in many cases the rates of increase in creatinine and cystatin C concentrations were approximately linear for the deaths (Fig. 1a and b) and the overall

rate of change (estimated by linear regression of all samples) was used to construct ROC curves. The dCr/dt and dCyC/dt in survivors were also estimated using linear regression for direct comparison to data from survivors. Of the 13 survivors, only four were found to have a positive gradient for dCr/dt that was statistically different to zero (data not shown). The gradients were much higher for deaths [medians 9.0 μmol/L/h (IQR 5.3–14.8) for deaths and 0.3 μmol/L/h CHIR-99021 manufacturer (IQR −0.3 to 3.3) for survivors; P = 0.002, Mann–Whitney test]. The ROC curve had an area of 0.93 (95% CI 0.83–1.04). The best dCr/dt cut-off was >4.3 μmol/L/h (sensitivity 100%, specificity 85% and likelihood ratio 7) ( Fig. 2a). Of the 11 survivors for which dCyC/dt results were available, only one trend line was found to have a positive gradient and to be statistically different to zero (data not shown). The gradients were again statistically greater for deaths [median

0.049 mg/L/h (IQR 0.017–0.074) for deaths and 0.004 mg/L/h (IQR −0.004 to 0.005) for survivors; P = 0.0022, Mann–Whitney test]. The dCyC/dt ROC curve had an area of 0.97 (95% CI 0.90–1.04) and the best cutoff was determined to be >0.009 mg/L/h (sensitivity 100%, find more specificity 91% and likelihood ratio 11) ( Fig. 2b). In one of these patients the dCr/dt and dCyC/dt exceeded values noted in deaths ( Fig. 1a) and the creatinine concentration fulfill criteria for acute renal failure. This patient was not predicted to die according to the admission paraquat concentration. This patient survived to hospital discharge without receiving haemodialysis, but was lost to follow up so it those is not known whether death occurred later. Excluding the two patients

discharged alive but unable to be found at follow-up (creatinine data available for both patients but cystatin C data only available in one) improved the predictive value of creatinine but did not substantially alter the results of this analysis for cystatin C. Specifically, dCr/dt ROC AUC = 0.96 (95% CI 0.87–1.05); best cutoff >4.3 mg/L/h (sensitivity 100%, specificity 91% and likelihood ratio 11), and dCyC/dt ROC AUC = 0.97 (95% CI 0.89–1.05); best cutoff >0.009 mg/L/h (sensitivity 100%, specificity 90% and likelihood ratio 10. However, as noted in Fig. 1a and b, the concentration of creatinine and cystatin C did not increase (or decrease) consistently in every patient. Therefore, dCr/dt and dCyC/dt values as determined by linear regression could vary depending on the time of sampling. To evaluate the minimum duration of sampling post-admission for assessing the dCr/dt or dCyC/dt, the rates of change from the time of admission to each subsequent blood sample for an individual patient were determined.