Patients Hospitalized in General Wards via the Emergency Department: Early Identification of Predisposing Factors for Death or Unexpected Intensive Care Unit Admission—A Historical Prospective

Table 3

Variables identified as potential predisposing factors for 15-day clinical worsening during hospitalization on regular wards, by multivariate logistic regression and by Cox proportional hazards models.

Predisposing factor	Logistic regression^a			Cox model^b
	Adjusted odd ratio	95% CI	𝑃-value	Adjusted hazard ratio	95% CI	𝑃-value
Unknown current medications	2.96	1.58–5.54	0.0007	3.15	1.72–5.76	0.0002
Oxygen administration in the ED	2.86	1.74–4.70	<0.0001	2.31	1.43–3.73	0.0007
Serum chloride below 100 mmol/L	2.32	1.45–3.72	0.0005	1.92	1.22–3.03	0.0053
Chronic hypertension	2.24	1.37–3.67	0.0013	1.75	1.09–2.80	0.0204
Presence of at least 3 SIRS items	2.05	1.19–3.54	0.0099	1.76	1.05–2.94	0.0339
Platelet count below 150 g/L	2.03	1.17–3.50	0.0113	1.73	1.03–2.92	0.0401
Male gender	1.92	1.15–3.13	0.0123	1.92	1.08–2.86	0.0247
Systolic arterial pressure below 110 mmHg	1.86	1.14–3.06	0.0138	1.97	1.23–3.16	0.0051
Use of psychoactive drug(s)	1.66	1.00–2.77	0.0499	1.70	1.03–2.81	0.0380
Solid unresolved cancer*	0.09	0.01–0.66	0.0178	0.09	0.01–0.68	0.0200

ED: emergency department; CI: confidence interval; SIRS: systemic inflammatory response syndrome (ref. bone).
aOverall model fit: log likelihood chi-square = 109 (𝑃<0.0001). bOverall model fit: log likelihood chi-square = 86 (𝑃<0.0001). ∗A 10th predictor in addition to the nine cited in the text was identified. The presence of unresolved cancer appeared as protective against target event occurrence in our cohort. However, the 95% CI of odd-ratio and hazard ratio were very large because only one target event occurred in the 291 unresolved cancer patients. This is due to the fact that we did not consider deaths that resulted from do-not-resuscitate orders as target events, whereas this is a frequent mode of death in cancer patients. As this could appear as counterintuitive, we did not cite the variable “unresolved cancer” in our results section. As this may create instability in our model, we reran the same analysis without incorporating the variable “unresolved cancer.” Both models (logistic and Cox regressions) kept a good overall fit (𝑃<0.0001) with the same covariates (except for “use of psychoactive drugs” that no longer reached statistical significance), and the logistic model kept a rather good AUROC value: 0.78 (0.77–0.80). In both analyses, the variables “unknown taken medications” and “oxygen therapy in the ED” remained the most powerful predictors (odd ratio 2.7 and 2.8, resp.; hazard ratio 2.8 and 2.3, resp.).