Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning

What it claims

SPAFT (October 2026 preprint) argues that PPG foundation models do not reliably beat general-purpose time-series models, even after full fine-tuning. Its fix is to prune the encoder first and then fine-tune only a small, importance-selected set of filters. The authors report that the best performance can come from fine-tuning about 1% of model parameters. The motivating analysis finds that up to 50% of filter outputs are nearly negligible on downstream tasks and up to 80% of intermediate filters stay inactive. In PaPaGei-P, 71% to 78% of filters have an activation probability below 50%. These patterns vary by dataset, which the authors read as task-specific subnetworks that full fine-tuning would disturb.

How it works

The method targets convolutional (ResNet-style) PPG models only. The authors give two reasons: CNN-based PPG models perform best in most cases, and most Transformer-based ones are not open source. The four models studied are AnyPPG, PaPaGei-P, PaPaGei-S and Pulse-PPG. The "nodes" pruned are convolutional filters, and a "block" is a ResNet layer.

flowchart TD
  A[Freeze encoder, fit linear/logistic head] --> B[Taylor importance per filter]
  C[Cosine graph of kernel weights per block] --> D[Redundancy sets per-block prune quota]
  B --> E[Prune: quota picks layers, importance picks filters]
  D --> E
  E --> F[Re-estimate importance on pruned model]
  F --> G[Fine-tune only top filters, budget split by score concentration]

1. Importance. With the encoder frozen, a linear regression or logistic head is fitted. Each filter gets a gate. The loss change from removing it is approximated by a second-order Taylor expansion, using per-sample gate gradients with an empirical Fisher diagonal. This costs one pass over the training set, and the scores are not recomputed during later pruning steps.
2. Redundancy. Each block becomes a graph. Nodes are normalized first-conv kernel vectors, linked when cosine similarity is at least a threshold τ. Redundancy decides how much each layer is pruned toward a global ratio ρ, and importance decides which filters go.
3. Adaptive fine-tuning. Importance is re-scored on the compact model. Trainable filters are allocated per layer according to where high-scoring units concentrate, and the remaining filters stay frozen.

The authors also explore personalized, per-individual training.

Evidence

The empirical sparsity analysis uses the BUT PPG and DeepBeat datasets. AnyPPG is the exception: all of its filters contribute at least 1.8% relative output norm, while the other three models have filters with near-zero output. Blocks with an average relative output norm below 1% are counted as unimportant.

The main experiments cover four public datasets plus a new smart-ring dataset collected at a large Asian hospital, across ten tasks. The headline win rates are 82.8% for regression and 87.5% for classification. The fetched text does not include the results tables, the baselines behind the win rates, or the metrics, so these figures cannot be checked here.

Limits

• The fetched text stops partway through the method section. The redundancy scoring details, results, ablations and personalization results are missing.
• The method applies only to CNN encoders. Transformer PPG models were not tested.
• No code, weights, licence or availability of the smart-ring dataset is stated.
• The pruning ratio ρ and the cosine threshold τ are hyperparameters, and the shown text gives no values for them.