<scp>PF</scp> ‐ <scp>LSTM</scp> : Leveraging Stochastic Feature Selection and Continuous and Discrete Respiratory Parameter for Robust Respiratory Waveform Analysis

Accurate prediction of respiratory parameters is crucial for ventilator‐assisted care. Existing methods rarely predict continuous and discrete parameters jointly and often fail to distinguish important from non‐important features, degrading precision. We propose PF‐LSTM, a time‐series prediction framework combining a pre‐trained language model (PubMedBERT), a stochastic‐gate FeatureSelector, and an LSTM. PF‐LSTM fuses patients' vital signs, semantic information, and temporal dynamics: PubMedBERT encodes semantic representations from vital‐sign records; the FeatureSelector identifies features most relevant to respiratory parameters; and the LSTM models sequential dependencies to improve prediction. We evaluate PF‐LSTM on respiratory parameter prediction and find it outperforms mainstream baselines across both continuous and discrete targets, with superior mean squared error (MSE), mean absolute error (MAE), R ‐squared ( R 2 ), F1 score, and recall. Reported values are 0.318 (MSE), 0.304 (MAE), 0.679 ( R 2 ), 0.329 (F1), and 0.297 (recall), respectively. These results indicate PF‐LSTM delivers accurate, efficient, and feature‐aware prediction for clinical decision support. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.