150 Chapter 7 F94 values were an assigned maximum/minimum value due to death/discharge. On day 8, 1948 out of 6079 (32.0%) known S/F94 values were an assigned maximum/ minimum value. A sensitivity analysis excluding these assigned values is in the supplementary material. An intermediate clinical outcome should have a strong association with a definitive outcome. Using 28-day mortality as the definitive outcome, and including S/F94 values on both day 0 and day 5 as covariates in a linear regression model, we found a strong inverse association between S/F94 on day 5 and mortality: an increased risk of mortality at day 28 is associated with a lower value of S/F94 on day 5 (Figure 2D). The OR for 28-day mortality is 0.25 (95% confidence interval 0.23-0.28), meaning that for a 1 unit increase in S/F94 on day 5, the odds of 28-day mortality decrease by 75%. We also compared S/F94 with a widely-used intermediate outcome, the WHO scale. Since this scale records clinical decisions about therapy that are, in part, determined by the severity of hypoxic lung disease, a close relationship was expected with S/ F94 (Figure 2C). The distributions were consistent between patients meeting the inclusion criteria (Figure 2C) and unselected patients (Supplementary Figure 5A). The distribution of S/F94 values between outcomes at day 28 for patients meeting the inclusion criteria is similar on day 0 and day 5 (Supplementary Figure 5B and Supplementary Figure 5C). As expected, when there are no criteria for supplemental oxygen in the first 3 days since admission (unselected patients, Supplementary Figure 5D and Supplementary Figure 5E), there is a relative increase of patients with high S/ 0 2000 4000 6000 8000 0.70 0.75 0.80 0.85 Treatment effect (predicted 28−day mortality relative risk ratio) Sample size Outcome measure WHO sustained 1−level improvement 28−day mortality WHO sustained 2−level improvement WHO day 5 S/F94 day 5 Protocolised S/F94 day 5 Figure 3. Comparison of the number of patients needed, including 95% confidence interval, for the different outcome measures, using treatment effects between 0.85 and 0.70. The bottom line shows predicted sample size required when using a protocolised S/F94 measurement, rather than an opportunistic measurement
RkJQdWJsaXNoZXIy MTk4NDMw