Fig 1.
Estimated densities for the five risk factors in the gapminder data are shown along the diagonal.
Below the diagonal, pairwise scatterplots of the data including linear regression (black lines) with 95% confidence intervals (shaded grey regions) are provided. Above the diagonal, the R2 values, indicating the percent of variation accounted for by the linear relationships between the data, are provided. Source: multiple sources for the individual variables aggregated by www.gapminder.org.
Fig 2.
Pairwise plots of the risk factor variables are again provided but with nonlinear smoothing spline fits.
The fits are shown as black lines with 95% confidence intervals as shaded grey regions. Above the diagonal, the estimated pairwise mutual information between the variables in bits are provided. These values indicated the average amount of shared information between the observations. All mutual information values were obtained using k-nearest neighbors type estimators (see S1) with k = 20.
Fig 3.
The dependence between the risk factor variables captured via correlation using R2 values are shown in the left image.
The right image shows the dependence using mutual information measured in bits. The mutual information estimates were obtained using k-nearest neighbors type estimators (see the appendix) with k = 20.
Table 1.
Summaries of the statistics pertaining to the age and hemodynamic variables.
Table 2.
Summaries of categorical covariates pertaining to ASA status, emergent or nonemergent operation, and risk of surgical procedure.
ASA Status is a numerical score from one to six of overall patient health from one (healthy patients) to five (critically ill patients). Cases with an ASA Code of five were excluded as there were too few such cases for reliable analysis.
Fig 4.
The distribution of patient I(HR,MAP) is shown for the cohorts of patients undergoing cardiac and noncardiac procedures.
We will consider a patient’s I(HR,MAP) to be large (small) if it is greater than (less than) approximately 4 bits (2.5 bits).
Fig 5.
Two examples of patient HR and MAP data are shown above.
(A) The patient’s HR remains roughly constant throughout the procedure. The variability in MAP does not have a clear association with HR which is consistent with a small I(HR, MAP) = 2.1 bits. (B) In this case, HR and MAP appear much more tightly coupled and exhibit similar dynamics throughout the procedure which is reflected in a large I(HR, MAP) = 4 bits. The correlation coefficient of HR and MAP, shown in the figures, were small and nearly equal which may wrongly suggest that HR and MAP were similarly dependent in these two cases.
Table 3.
The mean and 95% confidence intervals of the hazard ratio associated with each of the linear indicators in the CPH.
The outcome of interest in this study was the time to death were right censored for death occurring more than 365 from the surgical procedure.
Fig 6.
The hazard ratio from the CPH is shown as a function covariates involving HR and MAP.
Mean estimates are shown as solid lines with 95% confidence intervals as dashed lines. The most clinically relevant associations with mortality are elevated mean HR, decreased mean MAP, and decrease I(HR, MAP).
Fig 7.
The hazard ratio from the CPH is shown as a function of the number of HR/MAP measurements.
The mean estimate is shown as a solid line with 95% confidence intervals as dashed lines.
Table 4.
AUCS from ROC curves from the logistic regression models.
The outcome of interest in this model was death within 30 days of the surgical procedure.