Fig. 3: Latent space projections of patient representations by model input type. | Communications Medicine

Fig. 3: Latent space projections of patient representations by model input type.

From: Foundation models enable wearable signal screening for cardiovascular disease among people living with HIV

Fig. 3: Latent space projections of patient representations by model input type.

We visualised the separability of patient-level CVD likelihood using PCA, UMAP, and t-SNE projections on (a) routine clinical features, (b) NormWear embeddings, and (c) PaPaGei PPG embeddings. Compared to clinical and ECG-derived representations, PaPaGei embeddings showed markedly better class separation between CVD and non-CVD patients across all dimensionality reduction methods. Compared to clinical and ECG-derived representations, PaPaGei embeddings showed markedly better class separation between CVD and non-CVD participants across all dimensionality reduction methods. This suggests that PaPaGei captures richer physiological signals relevant to cardiovascular status from raw PPG with a frozen encoder (no fine-tuning).

Back to article page