Intradialytic hypotension (IDH) is a common complication of hemodialysis and is associated with treatment interruption, reduced patient tolerance, cardiovascular events, and hospitalization. Existing machine learning studies often use only limited dynamic dialysis information and provide insufficient feature interpretation. This study proposes an IDH risk prediction method based on multi-domain dynamic features and an improved Sparrow Search Algorithm (ISSA). After temporal alignment, outlier handling, and missing-value imputation, a candidate feature set was constructed from 12 baseline variables and 48 time-domain, frequency-domain, and time–frequency-domain features extracted from six continuous dialysis parameters. ISSA was then used to select a 13-dimensional feature subset comprising eight baseline and five dynamic features. Several machine learning models were evaluated before and after feature selection. Using a public hemodialysis monitoring dataset from a tertiary medical center, XGBoost achieved the best overall performance. Adding the five selected dynamic features increased ROC-AUC from 0.8355 to 0.8410 (ΔAUC = 0.0055, 95% CI: 0.0010–0.0100, p = 0.020), indicating a statistically significant incremental contribution. TreeSHAP analysis identified the spectral centroid of dialysis duration, conductivity, dialysate temperature, dialysis duration, and pre-dialysis body weight as important predictors. The proposed framework supports compact and interpretable early IDH risk prediction.