Research topic guides

Curated guides to PPG and wearable datasets and foundation models by task: blood pressure, atrial fibrillation, sleep staging, SpO2, stress, activity recognition, pretraining and more.

15 entries

• Best PPG datasets for blood pressure estimationCompare public PPG blood pressure datasets for machine learning: PulseDB, VitalDB, MIMIC-III-Ext-PPG, MIMIC-BP, UCI cuffless BP, PPG-BP and more. Subjects, sampling, BP reference method, licence and access.Dataset guide
• Wrist PPG datasets with raw waveformsMost smartwatches never export the PPG waveform, so public wrist datasets with raw signals are the main way to build and test wearable algorithms. This list only includes datasets where the BVP or PPG waveform itself is downloadable, with heart rate as an extra when present. Check the sampling rate and whether accelerometer data is included if you work on motion artifacts.Dataset guide
• PPG foundation models comparedA side-by-side view of foundation models that take PPG as input, from PaPaGei and Pulse-PPG to multimodal models such as NormWear and CSFM. Vision4PPG is a vision-foundation-model adaptation on imagified PPG, and CEReBrO is an EEG foundation model fine-tuned on ECG and PPG. FairTune is a bias-aware fine-tune of PPG-GPT for heart-rate fairness, not a separate pretraining run. None of those three is a PPG-native pretraining run. The table lists pretraining data, model size, whether weights are downloadable and the license, which are usually the first questions when choosing a model for a new task.Model guide
• Open-weight PPG models you can downloadOnly some PPG models publish their weights. These are the ones you can download and run today, with links to the checkpoint, the code and the license. Several are research-only, so the license column matters if you plan to ship a product.Model guide
• Best datasets for PPG atrial fibrillation detectionPublic PPG and ECG datasets for AFib detection with machine learning: MIMIC-III-Ext-PPG, MIMIC PERform AF, Pulsewatch, Chapman-Shaoxing and more. Labels, subjects, access and benchmarks.Dataset guide
• Best PPG datasets for heart rate estimation under motionCompare wrist PPG heart rate datasets with motion artefacts: PPG-DaLiA, WESAD, IEEE SPC 2015, WildPPG, Pulse Transit Time PPG and more. ECG reference, sampling, activities, licence and MAE benchmarks.Dataset guide
• Best wearable datasets for stress and emotion detectionCompare wearable stress and affect datasets: WESAD, UBFC-Phys, K-EmoCon, CLAS and more. Signals, protocol, labels, subjects, licence and AUROC benchmarks for PPG stress detection.Dataset guide
• Best wearable sleep staging datasetsPublic wearable sleep datasets with sleep-stage labels: DREAMT, Apple Watch sleep-accel, BIDSleep, PSGWatch, Newcastle PSG, EESM19, MESA, SHHS and Sleep-EDF, compared by signals and access.Dataset guide
• Best SpO2 and pulse oximetry datasets (and skin tone fairness)Public pulse oximetry and SpO2 datasets for machine learning and skin-tone bias research: OpenOximetry, BIDMC, Pulse Transit Time PPG, phone oximeter and more. Red/IR channels, references, access.Dataset guide
• Best public datasets for PPG foundation model pretrainingLarge public PPG, ECG and accelerometer corpora for self-supervised pretraining: VitalDB, MIMIC-III waveforms, MIMIC-III-Ext-PPG, PulseDB, MESA, NHANES and UK Biobank. Hours, subjects, licence and access.Dataset guide
• Best PPG and wearable datasets for glucose estimationPublic datasets pairing PPG or wearables with glucose: AI-READI, BIG IDEAs, BrisT1D, D1NAMO, VitalDB and fingerstick PPG sets, plus published MAE, MARD and Clarke zone A results and a mean-baseline reality check.Dataset guide
• Best wearable datasets for human activity recognitionPublic HAR datasets compared: CAPTURE-24, PAMAP2, UCI-HAR, OPPORTUNITY, WISDM, MHEALTH, USC-HAD, HHAR, RealWorld and HAR70+, with sensors, subjects, licences, Python guides and published macro-F1 results.Dataset guide
• PPG datasets with synchronised ECGDatasets that record PPG and ECG together let you use ECG R-peaks as the reference for heart rate and inter-beat intervals, measure pulse arrival time for blood pressure work, and test PPG beat detection against a gold standard. This list only includes datasets where both raw waveforms can be downloaded, sorted by number of subjects. Check whether the two signals share a clock, and the sampling rate of each, before you align beats.Dataset guide
• PPG datasets with a respiration referenceRespiratory rate can be estimated from PPG because breathing modulates the baseline, amplitude and frequency of the pulse wave. These datasets record PPG alongside a respiration reference, such as an impedance or belt signal or capnography, so estimates can be checked against measured breaths. They are sorted by number of subjects, with datasets that do not report a count listed last.Dataset guide
• Wearable datasets with PPG and EDAElectrodermal activity (EDA) tracks sympathetic arousal through changes in skin conductance. Combined with heart rate and heart rate variability from PPG, it is the usual input for wearable stress, emotion and workload models. These datasets record PPG and EDA together, sorted by number of subjects. Check the labelling protocol and whether the EDA signal is raw or already processed before comparing results across datasets.Dataset guide