TABLE 6

Comparison of various subsets of baseline characteristics to determine which provide the best model for prediction of change in visual analogue scale (VAS)

Model	Subset	Variables^#	AUC	Specificity	Sensitivity	PPV	NPV	Likelihood ratio
								Positive	Negative
1	Pre-drainage VAS	Pre-drainage VAS	0.80	0.61	0.87	0.86	0.63	2.23	0.21
2	All variables	Diagnosis Serum albumin Diaphragm movement pH Pre-drainage VAS	0.85	0.78	0.82	0.91	0.61	3.73	0.23
3	Demographics	None
4	Clinical outcomes	Total volume drained mL Time since detection days	0.62	0.56	0.61	0.79	0.34	1.39	0.70
5	Patient reported outcomes Demographics	Pre-drainage VAS	0.80	0.61	0.87	0.86	0.63	2.23	0.21
6	Patient reported outcomes (excluding pre-drainage VAS) Demographics	D12 physical	0.68	0.67	0.59	0.83	0.37	1.79	0.61
7	Radiology	Loculations	0.53	0.08	0.98	0.75	0.6	1.07	0.25
8	Physiology	FEV₁ % predicted	0.57	0.56	0.64	0.8	0.36	1.45	0.64
9	Physiology Radiology	Loculations	0.53	0.08	0.98	0.746	0.60	1.07	0.25
10	Demographics Clinical outcomes	Total volume drained mL Time since detection days	0.62	0.56	0.61	0.79	0.34	1.39	0.7
11	Demographics Clinical outcomes Patient reported outcomes (excluding pre-drainage VAS)	Total volume drained mL Serum albumin pH D12 physical	0.75	0.72	0.70	0.87	0.46	2.5	0.42
12	Demographics Clinical outcomes Patient reported outcomes (pre-drainage VAS, D12 and Borg)	Diagnosis Total volume drained mL Serum albumin pH Pre-drainage VAS	0.84	0.75	0.79	0.90	0.56	3.16	0.28
13	Demographics Clinical outcomes Physiology Radiology	Loculations	0.53	0.08	0.98	0.746	0.60	1.07	0.25
14	All variables (except pre-drainage VAS)	S_pO₂ Serum albumin Total volume drained mL pH D12 physical	0.76	0.78	0.70	0.90	0.48	3.18	0.38

Models were fitted using logistic regression and a change of 14 mm or more as a positive response. A model that included all variables provided the best predictive measures on the whole, with sensitivity=0.82, specificity=0.78, PPV=0.91, NPV=0.61 and AUC=0.85. However, all models were suboptimal and many other models provided results close to this. The prediction value was largely driven by the baseline VAS score, which by itself would provide similar predictive power (sensitivity=0.87, specificity=0.61, PPV=0.86, NPV=0.63 and AUC=0.80). All variables were pre-drainage with the exception of total volume drained. A value of n=135 for complete cases for all variables was used in order to compare between models. All variables considered in each of the 14 models are described in table E2. AUC: area under the curve; PPV: positive predictive value; NPV: negative predictive value; D12: dyspnoea-12; FEV₁: forced expiratory volume in 1 s; S_pO₂: oxygen saturation measured by pulse oximetry. ^#: as selected by Akaike's information criteria (AIC).