Clinical decision making under uncertainty: a bootstrapped counterfactual inference approach

Table 5 Policy Learning: Rewards (\(\mu \pm \sigma\)) of policies learned using inverse propensity scoring (IPS) formulation (10 simulations). Optdigits and Letter are two multiclass classification datasets from the UCI repository [30]. LR=Logistic Regression. NN=Neural Network

Dataset	Expert Policy/ Logging Policy	\(\hat{h}_0\) - NN
		Single NN	NN Ensemble		Adversarial
		IPS	IPS\(_{inv}\)	IPS\(_{avg}\)	IPS
UCI	OPTDIGITS (10 actions)	0.939 ± 0.013	0.921 ± 0.035	0.940 ± 0.012	0.942 ± 0.012
UCI	LETTER (26 actions)	0.429 ± 0.034	0.425 ± 0.032	0.448 ± 0.041	0.471 ± 0.027
Warfarin	LR (3 actions)	0.493 ± 0.040	0.506 ± 0.037	0.492 ± 0.040	0.515 ± 0.038
	LR (5 actions)	0.457 ± 0.034	0.469 ± 0.033	0.458 ± 0.032	0.471 ± 0.030
	PHARMA (3 actions)	0.656 ± 0.017	0.610 ± 0.028	0.640 ± 0.018	0.657 ± 0.019
	PHARMA (5 actions)	0.596 ± 0.020	0.516 ± 0.022	0.556 ± 0.020	0.626 ± 0.012
Heparin	Clinician (unknown)	0.295 ± 0.043	0.317 ± 0.033	0.311 ± 0.043	0.306 ± 0.035

ISSN: 1472-6947