Spatiotemporal Attention

The mechanism combines two dimensions of evidence rather than treating every observation as equally informative. It can assign higher weight to a tissue or lesion region while also emphasizing particular imaging frames or physiological-signal changes. This joint weighting helps the model connect a meaningful location with a meaningful time point, which is important when disease-related patterns evolve.

Spatiotemporal attention can capture relationships across locations and time points, not merely rank isolated regions or frames. A model may therefore relate a finding in one tissue area to how that area or another region changes later in a sequence. This is relevant to dynamic imaging and longitudinal records, where meaning depends on progression as well as position.

Attention weights matter because they indicate which regions and events contributed to a model’s output. In medical AI, those signals can provide clues about whether a prediction was influenced by a lesion, tissue area, imaging frame, or physiological change. This makes the mechanism useful not only for prediction, but also for examining the information the system considered most informative.

Relevant inputs include medical videos, dynamic imaging, longitudinal patient records, and time-dependent biosignals. These sources contain information whose significance may change across frames, examinations, or signal intervals. Applying the mechanism allows analysis to account for both where a pattern appears and when it appears, supporting more focused computational assessment of changing medical information.

A system considers spatial information, such as tissues or lesions, alongside temporal information, such as imaging-frame changes or physiological-signal changes. It then assigns different weights to these features and uses the resulting emphasis to capture relationships across locations and time points. The resulting representation can support disease detection, treatment monitoring, or clinical prediction.

It is relevant when the target outcome depends on change over time or on the location of a finding. Examples in the medical context include detecting disease, monitoring treatment, and making clinical predictions from dynamic imaging, medical videos, longitudinal records, or biosignals. The method is particularly useful when informative evidence is distributed across both anatomy and temporal behavior.