Table 3.

Model performance using single input types compared to all inputs

Single Input	Best DT	Optimized HPs	F1 (95% CI)	Precision (95% CI)	Recall (95% CI)
Medications	0.7	Dense units: 80, Learning rate: 0.000248, # LSTM layers: 2, LSTM units layer 1: 152, Dropout layer 1: 0.5, LSTM units layer 2: 112, Dropout layer 2: 0.5	0.217 (0.213– 0.220)	0.188 (0.185–0.191)	0.256 (0.252–0.260)
Laboratory	0.7	Dense units: 88, Learning rate: 1.518e-05, # LSTM layers: 1, LSTM units layer 1: 240, Dropout layer 1: 0.7	0.214 (0.211–0.218)	0.190 (0.186–0.194)	0.246 (0.241–0.250)
% Meals Consumed	0.5	Dense units: 64, Learning rate: 4.693e-05, # LSTM layers: 2, LSTM units layer 1: 32, Dropout layer 1: 0.8, LSTM units layer 2: 24, Dropout layer 2: 0.9	0.0688 (0.0671–0.0701)	0.0420 (0.0409–0.0429)	0.190 (0.186–0.194)
Diet Orders	0.5	Dense units: 64, Learning rate: 5.892e-05, # LSTM layers: 2, LSTM units layer 1: 120, Dropout layer 1: 0.8, LSTM units layer 2: 96, Dropout layer 2: 0.5	0.0926 (0.0912–0.0939)	0.0523 (0.0515–0.0531)	0.401 (0.395– 0.406)
Static History	0.6	Dense units: 72, Learning rate: 0.000451, # LSTM layers: 1, LSTM units layer 1: 56, Dropout layer 1: 0.6	0.153 (0.150–0.156)	0.112 (0.110–0.115)	0.239 (0.235–0.244)

Performance of LSTM models trained on individual input modalities. Input types included medications, laboratory values, % meals consumed, diet orders, and static patient history. For each model, the best-performing decision threshold (DT), optimized hyperparameters (HPs), and resulting mean F1 score, precision, and recall with 95% confidence intervals are shown.