Personalized Forecasting and Anomaly Detection for IoT Health Monitoring

Abstract

Reliable analytics for Internet of Things (IoT) health monitoring require temporal evaluation, correctly aligned anomaly scores and meaningful baselines. This study evaluates statistical and machine-learning methods using two fully synthetic data families. We first reproduce an existing comparison of univariate ARIMA and a multilayer perceptron (MLP) forecaster, and of Isolation Forest and an MLP autoencoder, on ten one-day simulated patients. Although the reported results reproduce, the anomaly experiment uses in-sample scoring and future observations, and averages window errors against point labels. Holding the fitted autoencoder unchanged and correcting only score alignment increases glucose F1 from 0.164 to 0.882. In a separate causal evaluation on 30 simulated glucose trajectories, mean F1 is 0.833 for Isolation Forest and 0.826 for the corrected autoencoder. A training-mean baseline reduces glucose RMSE from ARIMA’s 18.89 to 13.32 mg/dL on the original fixed-origin task. Further experiments use seven-day trajectories with daily structure, temporally correlated fluctuations and injected events. On 30 fresh trajectories, a patient-specific daily template with optional recent-residual correction achieves 30-minute RMSE of 3.37 mg/dL, compared with 4.37 for previous-day prediction and 5.75 for Kalman forecasting. Seasonal residual alerts with a separate sensor-quality check perform well on the tested perturbations; adding CUSUM does not improve recall for these scenarios. These findings support stronger evaluation and a simple personalized baseline, but do not establish clinical effectiveness or general superiority across model classes