Temporal heart-defect prediction (CheXchoNet)

Predicting heart defects from chest X-rays alongside the temporal metadata that comes with them, rather than treating each frame in isolation. A CNN trained over the CheXchoNet release reached AUC of 0.79 to 0.84 in cross-validation. The approach pointed me towards waveform-level signals, which is where the ICP work picks up.

Question

Most chest-X-ray classifiers treat each image as an isolated frame. Patients in the CheXchoNet release come with temporal metadata — repeat imaging, follow-ups, echo-confirmed diagnoses over time. The thesis asked whether a CNN that consumes imaging plus this temporal context outperforms a frame-only baseline at predicting congenital heart defects.

Approach

A CNN backbone over the chest-X-ray frames, with temporal metadata fused at a late stage (a small MLP merges per-patient temporal features with the imaging embedding). Training was over the CheXchoNet release with patient-level splits to prevent leakage.

Results

AUC of 0.79 to 0.84 across cross-validation folds. The temporal fusion gave a consistent improvement over the frame-only baseline, but the gap was smaller than I expected — the bigger lever turned out to be careful patient-level splitting and label cleaning, not the model architecture.

Where it pointed

The result that stuck with me was that signal-level data (the kind in raw waveforms) carries information that gets compressed away in image-level representations. That's the line of thinking the current ICP work picks up.