Table 1. Evaluating models on the ability to predict benzodiazepine usage from actigraphy.

In this dataset, the input is actigraphy and the label indicates whether a participant is taking benzodiazepines. Each model is trained on dataset sizes “500”, “1,000”, “2,500”, and “5,769” (seen in the columns) and evaluated using AUC on a held-out test set of 2,000 participants. The “Avg AUC” represents the averaged AUC scores across each training dataset size. If the model name has “smoothing” after it, it denotes that the model was trained on smoothed data. Underlined text indicates the best baseline model. PAT-S/M/L denotes Small, Medium, Large. A bolded PAT model indicates that it performed better than the best baseline, and a bolded and underlined PAT indicates the model with the best performance. PATs significantly outperform the baseline models in every dataset size in this task.

MODEL	Avg AUC*	n=500	n=1000	n=2500	n=5769	Params
LSTM	0.493	0.501	0.487	0.474	0.512	15 K
LSTM (smoothing)	0.499	0.506	0.508	0.482	0.499	15 K
Wavelet Transform	0.620	0.674	0.625	0.598	0.583	10 K
CNN-1D	0.632	0.621	0.630	0.640	0.637	10 K
CNN-1D (smoothing)	0.639	0.633	0.634	0.644	0.646	10 K
Conv LSTM (smoothing)	0.667	0.666	0.680	0.653	0.671	1.75 M
Conv LSTM	0.668	0.663	0.681	0.650	0.677	1.75 M
CNN-3D	0.693	0.683	0.693	0.693	0.703	790 K
CNN-3D (smoothing)	0.697	0.677	0.695	0.696	0.719	790 K
PAT-S	0.701	0.706	0.718	0.677	0.703	285 K
PAT Conv-S	0.726	0.737	0.711	0.722	0.735	285 K
PAT-M	0.744	0.743	0.745	0.742	0.745	1.00 M
PAT Conv-M	0.761	0.753	0.756	0.760	0.773	1.00 M
PAT Conv-L	0.762	0.763	0.756	0.754	0.773	1.99 M
PAT-L	0.767	0.771	0.765	0.760	0.771	1.99 M