Clinical Time Series Prediction Benefits from Missing Data as Signal

Clinical time series data, particularly from intensive care units, often presents a challenge due to incomplete measurements. This paper proposes that the very absence of data, or "missingness," can be a valuable indicator of clinical decisions and patient severity, thus serving as a predictive signal. The CISM framework addresses this by transforming each clinical variable into a time-frequency spectrogram. It then integrates an explicit "missingness stream" alongside this representation, preserving variable identity through aligned encoding. This approach allows the model to learn from both observed data patterns and the patterns of missing observations. Evaluations on an in-hospital mortality prediction task using the MIMIC-IV dataset demonstrated CISM's superior performance compared to existing baselines. Ablation studies confirmed that modeling observation patterns significantly contributes to predictive value, with the aligned missingness stream offering complementary gains in accuracy and precision.