Fig. 3: LAURA’s patient-dependent performance.
From: Transformer-based long-term predictor of subthalamic beta activity in Parkinson’s disease
LAURA performance on patient NWK1 assessed by monitoring the weighted mean average percentage error (wMAPE) between the observed and predicted distributions (see Suppl. Fig. 4 for all patients) as a similarity metric (Methods). a One-step-ahead prediction over a single aDBS setting. The prediction error is monitored for six different length (from 1 to 6 daily distributions) of the input sequence. The wMAPE of LAURA (dark red) visualized against the wMAPE of a deterministic linear regressor (red), collapsing to a zero-order regressor (light red) when the input sequence has length 1 sample. LAURA outperforms both the zero-order regressor when based on only 1-day of history, and the linear regressor when increasing the length of the history sequence. The zero and first-order regressors show a deterioration in forecasting performance when the length of the history sequence is increased from 1–3 daily distributions, and then an improvement in performance towards convergence with the results obtained with a history of only 1–2 days by increasing the length of the sequence. The forecasts of LAURA show quantitative and qualitative improvements independent of the length of the history sequence on which the forecast is based. b Multi-step-ahead prediction over a single aDBS setting. The wMAPE is monitored on predictions from 1 up to 6 days ahead, based on the patient’s history sequence of optimal length N*(=2 days, for patient NWK1) found experimentally. The wMAPE of LAURA (dark green) visualized against the wMAPE of a linear regressor (green) and the wMAPE of a zero-order regressor (light green). LAURA outperforms both zero and first-order regressors in multi-step-ahead prediction. While linear regressors increase the error in their predictions as the prediction step-ahead increases, LAURA manages to keep the error significantly lower than that computed by linear regressors and almost constant. c Multi-step-ahead prediction over multiple aDBS settings. LAURA can cope with intra-patient variability due to the recalibration of the aDBS device. The error of the predicted distributions is comparable to that obtained on one set of parameters.
