Table A2.

Optimal parameters (model, window, and feature selection correlation threshold) and number of features before and after training for all experiments. Simple imputation and feature selection with a correlation threshold were optimal for all experiments. Note that for each experiment, the feature counts correspond to the features available at the optimal window after hyperparameter tuning. The feature counts after training also correspond to the optimal feature selection method after hyperparameter tuning.

				Number of features before training		Number of features after training
	Model	Window (days)	Correlation threshold	Raw features	Total features (engineered)	Raw features	Total features (engineered)
UCLA Model	xgb	5	0.08	2350	8974	131	187
Cedars Sinai Model	lgb	6	0.05	1050	2989	212	281
UCLA + Cedars Sinai Model	lgb	10	0.02	3044	12379	1338	2838
UCLA + Cedars Sinai + Control Model	xgb	7	0.04	3238	13321	289	354
UCLA Model, Feature Intersection of All Cohorts	xgb	5	0.08	575	1024	61	80
UCLA Model, Feature Intersection of UCLA and Cedars Sinai	xgb	5	0.08	651	1121	60	80
Cedars Sinai Model, Feature Intersection of All Cohorts	lgb	6	0.05	595	1052	128	162
Cedars Sinai, Feature Intersection of UCLA and Cedars Sinai	xgb	14	0.035	703	1240	268	384
UCLA + Cedars Sinai Model, Feature Intersection of All Cohorts	lgb	10	0.02	635	1131	377	638
UCLA + Cedars Sinai Model, Feature Intersection of UCLA and Cedars Sinai	lgb	10	0.02	696	1218	411	687
UCLA + Cedars Sinai + Control Model, Feature Intersection of All Cohorts	lgb	10	0.01	635	1131	456	778