Deep PPG
Baseline in GPT-PPG: a GPT-based foundation model for photoplethysmography signals
7.47↓
Heart rate · WESAD · MAE (BPM)
leave-one-session-outwithin-datasetfine-tuned
Results explorer
One row per reported result from wearable PPG, ECG, motion and sleep papers. Filter by task, dataset, metric, sensor and year, or show only SOTA rows. Each value links to its source: a paper table or figure, or the leaderboard row it came from.
Last updated
433
reported results
224
with a table or figure reference
92
papers and models
134
task × dataset × metric pairs
38
benchmarks with a SOTA star
433 results
Heart rate · WildPPG · MAE (BPM)
subject-disjoint 5-fold CV, 5 seedswithin-datasetlinear probe on frozen embeddings
10.83↓
Heart rate · WildPPG · MAE (BPM)
subject-disjoint 5-fold CV, 5 seedswithin-datasetlinear probe on frozen embeddings
11.26↓
Heart rate · WildPPG · MAE (BPM)
subject-disjoint 5-fold CV, 5 seedswithin-datasetlinear probe on frozen embeddings
MOMENT
Baseline in PPG-Bench: A Unified Benchmark for PPG Foundation Models
11.54↓
Heart rate · WildPPG · MAE (BPM)
subject-disjoint 5-fold CV, 5 seedswithin-datasetlinear probe on frozen embeddings
PaPaGei-P
Baseline in PPG-Bench: A Unified Benchmark for PPG Foundation Models
11.96↓
Heart rate · WildPPG · MAE (BPM)
subject-disjoint 5-fold CV, 5 seedswithin-datasetlinear probe on frozen embeddings
12.63↓
Heart rate · WildPPG · MAE (BPM)
subject-disjoint 5-fold CV, 5 seedswithin-datasetlinear probe on frozen embeddings
0.86↑
Human activity recognition (IMU) · CAPTURE-24 · Cohen’s κ (e.g. sleep staging)
Protocol not structured
Leaderboard row
2026
SSL ResNet, UK Biobank (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
0.737 ± 0.087↑
Human activity recognition (IMU) · CAPTURE-24 · Cohen’s κ (e.g. sleep staging)
subject-disjointuses extra training data4 classes (paper Table 4)
Leaderboard row
ResNet from scratch (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
0.703 ± 0.092↑
Human activity recognition (IMU) · CAPTURE-24 · Cohen’s κ (e.g. sleep staging)
subject-disjoint4 classes (paper Table 4)
Leaderboard row
Random forest (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
0.683 ± 0.101↑
Human activity recognition (IMU) · CAPTURE-24 · Cohen’s κ (e.g. sleep staging)
subject-disjoint4 classes (paper Table 4)
Leaderboard row
0.818↑
Human activity recognition (IMU) · CAPTURE-24 · Macro-F1 on HAR activities
4-class intensity (Walmsley)
Leaderboard row
2026
SSL ResNet, UK Biobank (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
0.726 ± 0.093↑
Human activity recognition (IMU) · CAPTURE-24 · Macro-F1 on HAR activities
subject-disjointuses extra training data4 classes (paper Table 4)
Leaderboard row
ResNet from scratch (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
0.708 ± 0.094↑
Human activity recognition (IMU) · CAPTURE-24 · Macro-F1 on HAR activities
subject-disjoint4 classes (paper Table 4)
Leaderboard row
Random forest (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
0.694 ± 0.099↑
Human activity recognition (IMU) · CAPTURE-24 · Macro-F1 on HAR activities
subject-disjoint4 classes (paper Table 4)
Leaderboard row
Human activity recognition (IMU) · CAPTURE-24 · Macro-F1 on HAR activities
3-class intensity: sedentary / light / MVPA
Leaderboard row
2026
Diffusion-Based Synthetic Data Pretraining for E
Diffusion-Based Synthetic Data Pretraining for Enhancing Activity Recognition
90.6↑
Human activity recognition (IMU) · DEO · Balanced accuracy (HAR / class-imbalance)
uses extra training data
Leaderboard row
2026
Human activity recognition (IMU) · hapt · Macro-F1 on HAR activities
10 classes (4 activities + 6 directional postural transitions; ward-relevant proxy task)
Leaderboard row
2026
Human activity recognition (IMU) · har-6bench-avg · Macro-averaged F1
subject-disjointuses extra training datamixed (per-bench label sets)
Leaderboard row
2026
0.621 probe / 0.603 FT↑
Human activity recognition (IMU) · HHAR · Macro-F1 on HAR activities
subject-disjointuses extra training data6 classes
Leaderboard row
Random forest (Yang)
95.09↑
Human activity recognition (IMU) · MHEALTH · Accuracy
subject-disjoint
MCAE (Azadi)
99.58↑
Human activity recognition (IMU) · MHEALTH · F1 on HAR activities (averaging not stated)
subject-disjoint
Single-task baseline (Azadi)
99.23↑
Human activity recognition (IMU) · MHEALTH · F1 on HAR activities (averaging not stated)
subject-disjoint
CAE (Azadi)
97.33↑
Human activity recognition (IMU) · MHEALTH · F1 on HAR activities (averaging not stated)
subject-disjoint
Random forest (Yang)
94.49↑
Human activity recognition (IMU) · MHEALTH · F1 on HAR activities (averaging not stated)
subject-disjoint
CHEAR
CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition
88.62 (%)↑
Human activity recognition (IMU) · MHEALTH · Macro-F1 on HAR activities
subject-disjointas ZARA table (12 classes)
Leaderboard row
2026
Human activity recognition (IMU) · MHEALTH · Macro-F1 on HAR activities
subject-disjoint12 classes
Leaderboard row
OntoAR-HAR
OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition
84.6 ± 2.6 (%)↑
Human activity recognition (IMU) · MHEALTH · Macro-F1 on HAR activities
subject-disjointas ZARA table (12 classes)
Leaderboard row
2026
68.6 probe / 72.4 FT (%)↑
Human activity recognition (IMU) · MHEALTH · Macro-F1 on HAR activities
subject-disjointuses extra training data12 classes
Leaderboard row
Human activity recognition (IMU) · opentslm-har-cot · Macro-F1 on HAR activities
OpenTSLM HAR-CoT answer options
Leaderboard row
Human activity recognition (IMU) · OPPORTUNITY · Macro-F1 on HAR activities
subject-disjoint4 classes
Leaderboard row
OntoAR-HAR
OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition
91.5 ± 1.6 (%)↑
Human activity recognition (IMU) · OPPORTUNITY · Macro-F1 on HAR activities
subject-disjointas ZARA table (4 classes)
Leaderboard row
2026
CHEAR
CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition
85.13 (%)↑
Human activity recognition (IMU) · OPPORTUNITY · Macro-F1 on HAR activities
subject-disjointas ZARA table (4 classes)
Leaderboard row
2026
SSL ResNet, UK Biobank (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
59.5 ± 8.5 (%)↑
Human activity recognition (IMU) · OPPORTUNITY · Macro-F1 on HAR activities
subject-disjointuses extra training data4 classes (paper Table 4)
Leaderboard row
Random forest (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
41.6 ± 18.5 (%)↑
Human activity recognition (IMU) · OPPORTUNITY · Macro-F1 on HAR activities
subject-disjoint4 classes (paper Table 4)
Leaderboard row
ResNet from scratch (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
38.3 ± 12.4 (%)↑
Human activity recognition (IMU) · OPPORTUNITY · Macro-F1 on HAR activities
subject-disjoint4 classes (paper Table 4)
Leaderboard row
Random forest (Yang)
92.21↑
Human activity recognition (IMU) · PAMAP2 · Accuracy
subject-disjoint
MCAE (Azadi)
94.88↑
Human activity recognition (IMU) · PAMAP2 · F1 on HAR activities (averaging not stated)
subject-disjoint
Single-task baseline (Azadi)
92.66↑
Human activity recognition (IMU) · PAMAP2 · F1 on HAR activities (averaging not stated)
subject-disjoint
Random forest (Yang)
91.64↑
Human activity recognition (IMU) · PAMAP2 · F1 on HAR activities (averaging not stated)
subject-disjoint
CAE (Azadi)
91.0↑
Human activity recognition (IMU) · PAMAP2 · F1 on HAR activities (averaging not stated)
subject-disjoint
OntoAR-HAR
OntoAR-HAR: Ontology-Guided Agentic RAG with Personalized Memory for Human Activity Recognition
87.3 ± 1.1 (%)↑
Human activity recognition (IMU) · PAMAP2 · Macro-F1 on HAR activities
subject-disjointas ZARA table (12 classes)
Leaderboard row
2026
CHEAR
CHEAR: Confusion-Hierarchy Evidence-Aware Reasoning for Cost-Efficient Human Activity Recognition
83.43 (%)↑
Human activity recognition (IMU) · PAMAP2 · Macro-F1 on HAR activities
subject-disjointas ZARA table (12 classes)
Leaderboard row
2026
SSL ResNet, UK Biobank (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
78.9 ± 5.4 (%)↑
Human activity recognition (IMU) · PAMAP2 · Macro-F1 on HAR activities
subject-disjointuses extra training data8 classes (paper Table 4)
Leaderboard row
Human activity recognition (IMU) · PAMAP2 · Macro-F1 on HAR activities
subject-disjoint12 classes
Leaderboard row
Random forest (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
75.3 ± 9.3 (%)↑
Human activity recognition (IMU) · PAMAP2 · Macro-F1 on HAR activities
subject-disjoint8 classes (paper Table 4)
Leaderboard row
75.15 LP / 75.80 FT (%)↑
Human activity recognition (IMU) · PAMAP2 · Macro-F1 on HAR activities
subject-disjointPAMAP2 activities incl. NULL (class count not stated in the source row)
Leaderboard row
2026
ResNet from scratch (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
60.5 ± 8.6 (%)↑
Human activity recognition (IMU) · PAMAP2 · Macro-F1 on HAR activities
subject-disjoint8 classes (paper Table 4)
Leaderboard row
Take it Personally
Take it Personally: The Limits of General SSL Representations for Real-Life PPG Emotion Detection
0.64↑
Human activity recognition (IMU) · prosi · Macro-F1 on HAR activities
subject-disjoint7 activities
Leaderboard row
2026
SSL ResNet, UK Biobank (Yuan)
Self-supervised learning for human activity recognition using 700,000 person-days of wearable data
0.792 ± 0.075↑
Human activity recognition (IMU) · RealWorld HAR · Macro-F1 on HAR activities
subject-disjointuses extra training data8 classes (paper Table 4)
Leaderboard row
Showing 251–300 of 433 results
Recent changes
New entry: A cross-subject stress recognition method…, 0.9158
Stress detection (acute / prolonged / affect): WESAD: accuracy
New entry: PC-AVCT wearable-state model, 7.697
From the hub's dated sync history: new results as they are added, and SOTA stars as they are awarded or move.
How SOTA is decided
- marks the best value for one task, dataset and metric, using the metric's direction (higher or lower is better).
- Results are compared only under the same evaluation protocol: split type, within- or cross-dataset testing, filtering, inputs and label set. Different protocols stay separate and are never ranked together.
- Only results with a table, figure or section reference compete, and a star needs at least two competing methods. Ties go to the earlier publication.
- Stars are recomputed whenever the research hub syncs, and every change is logged.
Downloads of every result (JSON and CSV) are free for members. Sign in to download.
Results as reported by the authors; not independently reproduced. Leaderboard rows without a structured protocol are listed but don't receive stars. See also the leaderboards.
Cite or reuse this
CC BY 4.0ChatPPG compiled this data. You can copy, adapt and republish it, including in articles, slides, papers and AI tools, as long as you credit ChatPPG and link to this page. Please also cite the original studies for any number you use.
Attribution
Source: ChatPPG, Wearable AI results explorer, https://chatppg.com/research/results (CC BY 4.0)
HTML link
Source: <a href="https://chatppg.com/research/results">ChatPPG, Wearable AI results explorer</a> (<a href="https://creativecommons.org/licenses/by/4.0/" rel="license">CC BY 4.0</a>)