Your smartwatch tracks your heart. Your EEG captures your brain. These devices provide better ways to understand stress, sleep, heart rhythm, and other aspects of one’s health. However, the bio-signals generated from these devices are noisy, differing from person to person. Teaching AI to make sense of these bio-signals takes a lot of labeled data and medical expertise to do it.

That’s when our researchers at Dolby came up with CiTrus—a convolution-transformer hybrid model built for low-data bio-signals transfer learning. The idea was simple: pre-train the model on a large EEG dataset, then fine-tune it on entirely different signals — like ECG, EMG, or PPG — using only a fraction of the labeled examples typically required.

Tested across eight bio-signal datasets, CiTrus consistently outperformed existing models, with performance gains of up to 60% on some datasets. They also report that multimodal pre-training can boost performance again and that their new dataset-alignment method markedly improves results when signal lengths and sampling rates differ.

Teaching Machines to Read Your Body: Dolby’s CiTrus Model

CiTrus goes further with two key innovations:

First, a frequency-based pre-training approach that teaches the model to understand both the raw signal and its frequency patterns simultaneously.

Second, a smarter fine-tuning technique that resamples new datasets to match the original training frequency, preserving the patterns the model already learned.

As bio-signals become more common in wearables and health technology, AI models will need to learn more from less data and to transfer that learning across very different signals. Dolby’s CiTrus points to a practical path forward where AI health tools that need minimal data to get started make intelligent bio-signals analysis faster, cheaper, and accessible at scale.

CiTrus: How Machines Think Smarter with Bio-signal AI with Less Data.

{{ pageEducationalSlot.headingContent[index] }}