Prediction of health care expenditure increase: how does pharmacotherapy contribute?

Table 2 Comparison of prediction performance of logistic regression (LR), boosted decision tree (BDT) and feedforward neural network (FNN) using different sets of features

		Models
*Model performance on validation dataset*		LR		BDT		FNN
Features	Size	Acc (%)	AUC	Acc (%)	AUC	Acc (%)	AUC
Demographic model*	7	51.2	0.52	51.3	0.52	52.2	0.53
+ number of different drugs	8	58.0	0.61	58.1	0.61	58.7	0.61
+ number of individual prescriptions	8	55.3	0.58	56.9	0.60	57.5	0.60
+ number of hospitalisations	8	61.0	0.62	61.0	0.62	61.1	0.63
+ number of outpatient physician office visits	8	59.4	0.63	60.1	0.63	60.4	0.64
+ chronic conditions	29	54.8	0.57	57.0	0.59	57.5	0.60
Extended model^†	33	62.8	0.67	63.1	0.68	64.0	0.69
+ additional features	297	64.8	0.70	66.3	0.72	66.1	0.72
+ features representing pharmacotherapy	482	64.5	0.69	65.4	0.71	65.6	0.71
+ total costs	34	62.1	0.67	64.8	0.71	65.7	0.71
+ additional features + total costs	298	65.0	0.70	67.0	0.74	67.0	0.73
Complete model^‡ without total costs	746	65.3	0.71	66.5	0.73	66.5	0.72
Complete model	747	65.2	0.70	67.4	0.74	67.4	0.73
Backward Deletion	36	–	–	66.9	0.73	–	–
*Model performance on test dataset*
Complete model^‡ without total costs	746	65.9	0.71	66.8	0.73	66.4	0.72
Complete model	747	65.7	0.71	67.6	0.74	67.2	0.73
Backward Deletion	36	–	–	67.1	0.73	–	–

Acc = Accuracy, AUC = Area under the curve, Size = Number of features in the model
^*Demographic model = age + gender + area of residence + deductible + insurance model,
^†Extended model = Demographic model + number of different drugs + number of individual prescriptions + number of hospitalisations + number of outpatient physician office visits + chronic conditions
^‡Complete model = Extended model + additional predictors + features representing pharmacotherapy + total costs
Bold data are significant

ISSN: 1472-6963