A Statistical Approach for Modeling Irregular Multivariate Time Series with Missing Observations

Irregular multivariate time series with missing values present significant challenges for predictive modeling in domains such as healthcare. While deep learning approaches often focus on temporal interpolation or complex architectures to handle irregularities, we propose a simpler yet effective alternative: extracting time-agnostic summary statistics to eliminate the temporal axis. Our method computes four key features per variable-mean and standard deviation of observed values, as well as the mean and variability of changes between consecutive observations to create a fixed-dimensional representation. These features are then utilized with standard classifiers, such as logistic regression and XGBoost. Evaluated on four biomedical datasets (PhysioNet Challenge 2012, 2019, PAMAP2, and MIMIC-III), our approach achieves state-of-the-art performance, surpassing recent transformer and graph-based models by 0.5-1.7% in AUROC/AUPRC and 1.1-1.7% in accuracy/F1-score, while reducing computational complexity. Ablation studies demonstrate that feature extraction-not classifier choice-drives performance gains, and our summary statistics outperform raw/imputed input in most benchmarks. In particular, we identify scenarios where missing patterns themselves encode predictive signals, as in sepsis prediction (PhysioNet, 2019), where missing indicators alone can achieve 94.2% AUROC with XGBoost, only 1.6% lower than using original raw data as input. Our results challenge the necessity of complex temporal modeling when task objectives permit time-agnostic representations, providing an efficient and interpretable solution for irregular time series classification.

💡 Research Summary

The paper tackles the problem of irregular multivariate time series with extensive missing values, a common situation in biomedical domains such as intensive care monitoring. While recent deep‑learning approaches (GRU‑D, SeFT, mTAND, Raindrop, ViTST, etc.) attempt to model the temporal dynamics directly, they often require sophisticated architectures, large numbers of parameters, GPU‑accelerated training, and careful handling of missingness. The authors propose a radically simpler pipeline that discards the temporal axis altogether and replaces each variable with four summary statistics: (1) mean of observed values, (2) standard deviation of observed values, (3) mean of successive differences (i.e., average change), and (4) standard deviation of those differences. When a variable has no observations in a segment, the global mean computed over the whole training set is used and the variance is set to zero; when there are no consecutive observations, the change‑related statistics are set to zero. By concatenating these four numbers for every variable, a segment of length L×D is transformed into a fixed‑size 4×D feature matrix, which is then flattened.

The extracted feature matrix is fed to conventional classifiers. The authors evaluate logistic regression, random forest, support vector machine, and especially XGBoost (100 trees, max depth 5, learning rate 0.1) using a uniform 5‑fold cross‑validation protocol. They benchmark on four publicly available biomedical datasets: PhysioNet Challenge 2019 (sepsis prediction, 34 variables, 40 k segments, >90 % missing for most variables), PhysioNet Challenge 2012 (in‑hospital mortality, 37 variables, 12 k segments, high missingness), PAMAP2 (human activity recognition, 17 inertial variables, 5 k segments, artificially 60 % missing), and MIMIC‑III (in‑hospital mortality, 17 variables, 21 k segments, >80 % missing). These datasets differ in sequence length, sampling regularity, and class imbalance, providing a rigorous test bed.

Results show that the XGBoost model built on the statistical summary consistently outperforms the state‑of‑the‑art deep‑learning baselines reported in the literature. Gains range from 0.5 to 1.7 percentage points in AUROC/AUPRC for binary tasks and 1.1 to 1.7 points in accuracy/F1 for the multi‑class activity task. Notably, using only the missing‑indicator mask as input yields an AUROC of 94.2 % on the sepsis task, demonstrating that the pattern of missingness itself carries strong predictive information. Ablation experiments confirm that the performance boost originates from the feature extraction step rather than the choice of classifier: feeding raw or imputed time series to the same XGBoost model results in lower scores than using the proposed summary features.

Beyond predictive performance, the method dramatically reduces computational overhead. No GPU is required; training and inference complete within seconds on a standard CPU, and memory consumption is minimal compared with transformer‑based or graph‑neural‑network models that involve millions of parameters. Interpretability is also enhanced: each feature corresponds to a transparent statistic (mean, variance, change) for a specific physiological variable, allowing clinicians to trace model decisions back to clinically meaningful patterns.

In summary, the study demonstrates that for many irregular time‑series classification problems—especially when the task does not demand fine‑grained temporal resolution—a time‑agnostic statistical representation can be both more accurate and far more efficient than complex deep‑learning architectures. The work challenges the prevailing assumption that sophisticated temporal modeling is necessary for irregular biomedical data and provides a practical, reproducible alternative that can be readily adopted in real‑world clinical decision support systems. Future directions may include enriching the feature set with non‑linear transformations, domain‑specific derived metrics, or hybrid models that combine the statistical summary with lightweight temporal encoders for tasks that do require some temporal nuance.