Benchmarking Generalist and Specialist Time-Series Foundation Models for PPG-Based Health Assessment
Organizations: Nell Hodgson Woodruff School of Nursing, Emory University. · School of Computer Science, University of Oklahoma. · Department of Computer Science, Emory University. · Department of Pediatrics, Emory University School of Medicine. ·
Abstract
Time-series Foundation Models (FMs) are pretrained on either heterogeneous corpora or specific signal domains. Using Photoplethysmography (PPG) as a representative modality, we compare generalist time-series FMs, PPG-specialist FMs, and traditional machine learning (ML) baselines for PPG-based health assessment. We conduct a two-stage benchmark. First, we evaluate multiple FM families across 23 downstream evaluations (e.g., atrial fibrillation detection, vital-sign estimation). Second, we conduct a focused four-FM comparison (MOMENT, Chronos-2, GPT-PPG, MOMENT-PPG) at multiple model sizes on an expanded 51-task suite, under both frozen-backbone and full-model tuning, to study tuning strategy, model scaling, and cross-dataset transfer. The multi-model benchmark shows that PPG-specialist FMs achieve the strongest aggregate performance, but their advantage is not uniform: generalist FMs remain competitive on several tasks, and traditional ML baselines outperform FMs on selected clinical regression tasks. The focused analysis shows that the advantage of PPG-specific pretraining appears mainly with frozen backbones and on targets that the waveform directly reflects, such as atrial fibrillation and vital signs. Full-model tuning narrows this advantage, model rankings change with the tuning strategy, and the largest model variant is the best choice in only about one-third of cases. Additionally, controlled experiments with two generalist models show that initializing PPG pretraining from generalist checkpoints provides an early advantage over training from scratch, even when the pretrained features offer little downstream benefit. Together, these findings suggest that PPG-specific pretraining is useful but task-dependent, and that FMs should be evaluated against simpler statistical, traditional ML, and clinical history-based baselines.
Figures & tables
| Model | Pretraining corpus | Architecture | Objective |
| Generalist time-series FMs | |||
| Chronos-2 | Public + synthetic time series | T5-family transformer | Multi-step quantile forecasting |
| TimesFM | 100B web/traffic time points | Decoder-only transformer | Next-patch prediction |
| Moirai-2 | Tens of millions of series | Decoder-only transformer | Quantile multi-token prediction |
| MOMENT | Time Series Pile, multi-domain | Transformer encoder | Masked time-series modeling |
| PPG-specialist FMs | |||
| Task | Dataset | Metric | Best generalist FM | Best PPG-specialist FM | Best traditional ML |
| AF Detection | Stanford | F1 (%) | Chronos-2: 87.77 0.85 | PulsePPG: 93.93 0.43 | CNN1D: 86.49 0.90 |
| AF Detection | Simband | F1 (%) | Chronos-2: 69.66 3.78 | GPT-PPG-19M: 67.91 3.84 | LogReg: 68.17 3.56 |
| Mental Workload | MAUS | F1 (%) | Chronos-2: 68.54 2.00 | PulsePPG: 72.75 2.13 | LSTM: 67.20 2.26 |
| Activity Recognition | Pulse Transit | Acc. (%) | MOMENT: 87.96 1.06 | PulsePPG: 87.46 0.94 | CNN1D: 84.16 0.99 |
| Heart Rate Estimation | WESAD | MAE | Chronos-2: 9.60 0.07 | PulsePPG: 9.93 0.08 | CNN1D: 9.94 0.07 |
| Heart Rate Estimation | DALIA | MAE | Chronos-2: 9.03 0.10 | PulsePPG: 5.91 0.09 | XGBoost: 24.25 0.13 |
| Model family | Model / baseline | Win score | Main pattern |
| Generalist FM | Chronos-2 | 3.5/23 | Highest generalist win count; competitive on several regression tasks |
| Generalist FM | TimesFM | 0/23 | Competitive on selected tasks but no task-level win |
| Generalist FM | Moirai-2 | 1/23 | One task-level win; otherwise mixed performance |
| Generalist FM | MOMENT | 3/23 | Strong on selected classification and clinical tasks |
| PPG-specialist FM | PaPaGei | 0.5/23 | Shares the best SpO 2 result on MIMIC-III; otherwise no task-level win |
| PPG-specialist FM | PulsePPG | 8/23 | Highest overall win count, especially on PPG-centered tasks |
| Token mixer | Type | Tasks | Median gain | Composite |
| Local convolution | Convolution | 14/16 | ||
| LRU | Recurrent | 13/16 | ||
| Selective state-space | State-space | 13/16 | ||
| Hyena | Convolution | 12/16 | ||
| Global convolution | Convolution | 14/16 | ||
| GRU | Recurrent | 11/16 |
| Token mixer | Pre-training objective | ||
| \ltx@saved@cline 2-4 | Masked | Next-patch | Contrastive (control) |
| Local convolution | 11/16 ( ) | 13/16 ( ) | 6/16 ( ) |
| LRU | 13/16 ( ) | 14/16 ( ) | 8/16 ( ) |
| Selective state-space | 14/16 ( ) | 14/16 ( ) | 13/16 ( ) |
| Task baseline type | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG | Signal/morph. baseline |
| Heart Rate Estimation (MAE) Inter-Beat Interval (IBI) | |||||
| WESAD | 10.87 ± 0.09 | 9.78 ± 0.07 | 10.84 ± 0.08 | 9.54 ± 0.07 | 19.52 0.15 |
| DALIA | 12.98 ± 0.11 | 9.30 ± 0.11 | 9.39 ± 0.09 | 7.01 ± 0.08 | 16.80 0.14 |
| Respiration Rate Estimation (MAE) Baseline Wander (BW) | |||||
| BIDMC | 2.09 ± 0.06 | 1.57 ± 0.05 | 1.30 ± 0.05 | 1.40 ± 0.05 | 5.85 0.56 |
| MIMIC-III | 4.64 ± 0.04 | 4.45 ± 0.04 | 4.50 ± 0.04 | 4.38 ± 0.04 | 8.04 0.19 |
Appendix figures & tables32 assets
Supplementary material from the paper’s appendix.
Appendix
| Group | # | Features |
| Time-domain statistics | 13 | mean, standard deviation, minimum, maximum, median, interquartile range, skewness, kurtosis, root mean square, zero-crossing rate, mean absolute deviation, range, energy |
| Pulse morphology | 11 | peak count, peak-to-peak amplitude, inter-beat interval (IBI) mean, IBI standard deviation, IBI RMSSD, pNN50, mean rise time, mean fall time, mean maximum upstroke slope, mean pulse area, mean pulse width |
| Spectral | 8 | dominant frequency, dominant power ratio, spectral entropy, VLF power (0–0.5 Hz), LF power (0.5–4 Hz), HF power (4–10 Hz), total power, signal-to-noise ratio |
| Autocorrelation and complexity | 5 | autocorrelation at lag 1, autocorrelation at lag 10, Hjorth mobility, Hjorth complexity, sample-entropy proxy |
| Task | Task Type ( ) | Data Source | Channels | #participants | Data size | Length |
| AF Detection | Classification (2) | Stanford Torres-Soto and Ashley (2020) | 1 | 163 | 285 h | 25 s |
| \cline 3-7 | Simband Shashikumar et al. (2017) | 1 | 98 | 7 h | 30 s | |
| Mental Load Assessment | Classification (2) | Mental Stress PPG Anwar and Zakir (2022) | 1 | 27 | 4.77 h | 60 s |
| \cline 3-7 | MAUS Beh et al. (2021) | 1 | 22 | 4.375 h | 5 s | |
| Activity Recognition | Classification (3) | Pulse Transit PPG Mehrgardt et al. (2022) | 2 | 22 | 9 h | 5 s |
| Human Identification | Classification (35) | Real World PPG Siam et al. (2019) | 1 | 35 | 3.5 h | 6 s |
| Task (Dataset) | Chronos-2 (120M) | TimesFM | Moirai-2 | MOMENT (125M) | PaPaGei |
| Classification (F1 / accuracy %, higher is better) | |||||
| AF Detection (Stanford) | 87.77 0.85 | 80.25 1.14 | 83.51 1.00 | 83.55 1.09 | 81.43 0.77 |
| AF Detection (Simband) | 69.66 3.78 | 60.11 4.19 | 54.76 4.28 | 55.22 3.94 | 61.75 2.91 |
| Mental Workload (MAUS) | 68.54 2.00 | 66.19 2.18 | 62.45 2.34 | 68.01 2.37 | 58.42 2.40 |
| Mental Stress (Mental Stress PPG) | 68.10 5.71 | 70.11 5.75 | 70.06 5.52 | 72.73 5.49 | 55.95 7.33 |
| Activity Recognition (Pulse Transit) | 77.72 1.14 | 85.85 0.88 | 78.00 1.13 | 87.96 1.06 | 67.39 1.23 |
| Task (Dataset) | PulsePPG | GPT-PPG-19M | GPT-PPG-85M | Best traditional ML |
| Classification (F1 / accuracy %, higher is better) | ||||
| AF Detection (Stanford) | 93.93 0.43 | 86.99 0.91 | 87.63 0.84 | CNN1D: 86.49 0.90 |
| AF Detection (Simband) | 60.60 3.12 | 67.91 3.84 | 66.01 3.82 | LogReg: 68.17 3.56 |
| Mental Workload (MAUS) | 72.75 2.13 | 61.57 2.83 | 64.32 2.58 | LSTM: 67.20 2.26 |
| Mental Stress (Mental Stress PPG) | 65.59 6.34 | 71.11 5.55 | 64.86 6.61 | CNN1D: 70.74 5.43 |
| Activity Recognition (Pulse Transit) | 87.46 0.94 | 78.81 1.12 | 80.37 1.07 | CNN1D: 84.16 0.99 |
| Task/Dataset (Metric) | MOMENT | Chronos-2 | |||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-6 | 40M | 125M | 385M | 28M | 120M |
| AF Detection (F1 Score, %) | |||||
| Stanford | 55.47 1.55 | 61.20 1.45 | 66.74 1.54 | 62.95 1.40 | 66.77 1.43 |
| Simband | 59.74 5.06 | 58.19 4.36 | 58.33 4.07 | 59.76 4.15 | 57.58 4.35 |
| Mental Workload (F1 Score, %) | |||||
| MAUS | 65.96 2.44 | 62.90 2.47 | 72.72 2.56 | 60.38 2.62 | 64.12 2.28 |
| Task/Dataset (Metric) | GPT-PPG | MOMENT-PPG | ||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-7 | 19M | 85M | 345M | 40M | 125M | 385M |
| AF Detection (F1 Score, %) | ||||||
| Stanford | 84.34 0.95 | 85.83 0.95 | 62.80 1.31 | 68.36 1.28 | 74.53 1.27 | 75.89 1.18 |
| Simband | 58.95 3.89 | 64.89 3.57 | 52.00 4.46 | 53.10 4.09 | 60.34 3.77 | 60.50 3.75 |
| Mental Workload (F1 Score, %) | ||||||
| MAUS | 57.92 2.81 | 54.92 2.78 | 60.04 2.72 | 64.79 2.61 | 59.73 2.70 | 62.58 2.65 |
| Task/Dataset (Metric) | MOMENT | Chronos-2 | |||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-6 | 40M | 125M | 385M | 28M | 120M |
| AF Detection (F1 Score, %) | |||||
| Stanford | 85.59 1.09 | 83.55 1.09 | 89.63 1.11 | 82.30 1.06 | 87.77 0.85 |
| Simband | 64.32 4.34 | 55.22 3.94 | 64.84 3.72 | 67.45 3.87 | 69.66 3.78 |
| Mental Workload (F1 Score, %) | |||||
| MAUS | 68.24 2.35 | 68.01 2.37 | 66.03 2.19 | 68.22 2.05 | 68.54 2.00 |
| Task/Dataset (Metric) | GPT-PPG | MOMENT-PPG | ||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-7 | 19M | 85M | 345M | 40M | 125M | 385M |
| AF Detection (F1 Score, %) | ||||||
| Stanford | 86.99 0.91 | 87.63 0.84 | 87.25 0.88 | 85.89 0.92 | 89.83 0.81 | 91.02 0.72 |
| Simband | 67.91 3.84 | 66.01 3.82 | 61.11 4.04 | 50.21 3.89 | 61.54 3.95 | 68.77 3.40 |
| Mental Workload (F1 Score, %) | ||||||
| MAUS | 61.57 2.83 | 64.32 2.58 | 60.91 2.80 | 67.53 2.20 | 51.17 3.11 | 62.30 2.68 |
| Task/Dataset (MAE) | MOMENT | Chronos-2 | |||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-6 | 40M | 125M | 385M | 28M | 120M |
| SpO 2 Estimation (MAE) | |||||
| Sleep Disorder | 3.68 0.05 | 3.53 0.05 | 5.42 0.05 | 5.26 0.05 | 5.17 0.05 |
| BIDMC | 5.07 0.07 | 3.54 0.07 | 3.64 0.07 | 2.40 0.06 | 2.44 0.07 |
| MIMIC-III | 2.56 0.02 | 2.57 0.02 | 2.59 0.02 | 2.21 0.02 | 2.21 0.02 |
| Skin Temperature Estimation (MAE) | |||||
| Task/Dataset (MAE) | GPT-PPG | MOMENT-PPG | ||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-7 | 19M | 85M | 345M | 40M | 125M | 385M |
| SpO 2 Estimation (MAE) | ||||||
| Sleep Disorder | 6.67 0.08 | 4.04 0.08 | 4.05 0.08 | 3.93 0.07 | 4.14 0.07 | 3.98 0.07 |
| BIDMC | 2.04 0.07 | 2.58 0.08 | 2.44 0.07 | 2.07 0.06 | 1.92 0.06 | 1.94 0.07 |
| MIMIC-III | 2.83 0.03 | 2.29 0.03 | 2.46 0.03 | 2.19 0.02 | 2.18 0.02 | 2.20 0.02 |
| Skin Temperature Estimation (MAE) | ||||||
| Task/Dataset (MAE) | MOMENT | Chronos-2 | |||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-6 | 40M | 125M | 385M | 28M | 120M |
| Potassium Estimation (MAE) | |||||
| MIMIC-III | 0.49 0.004 | 0.48 0.004 | 0.50 0.004 | 0.46 0.004 | 0.46 0.004 |
| Inst. A | 0.46 0.01 | 0.46 0.01 | 0.45 0.01 | 0.47 0.01 | 0.48 0.01 |
| Sodium Estimation (MAE) | |||||
| MIMIC-III | 5.17 0.03 | 5.84 0.03 | 5.73 0.03 | 3.70 0.03 | 3.68 0.03 |
| Task/Dataset (MAE) | GPT-PPG | MOMENT-PPG | ||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-7 | 19M | 85M | 345M | 40M | 125M | 385M |
| Potassium Estimation (MAE) | ||||||
| MIMIC-III | 0.47 0.00 | 0.47 0.00 | 0.46 0.00 | 0.46 0.00 | 0.46 0.00 | 0.46 0.00 |
| Inst. A | 0.57 0.01 | 0.58 0.01 | 0.48 0.01 | 0.48 0.01 | 0.50 0.01 | 0.49 0.01 |
| Sodium Estimation (MAE) | ||||||
| MIMIC-III | 4.80 0.04 | 3.78 0.03 | 3.74 0.03 | 3.77 0.04 | 3.78 0.03 | 3.81 0.03 |
| Task/Dataset (MAE) | MOMENT | Chronos-2 | |||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-6 | 40M | 125M | 385M | 28M | 120M |
| SpO 2 Estimation (MAE) | |||||
| Sleep Disorder | 3.94 0.10 | 3.44 0.14 | 4.28 0.20 | 3.77 0.05 | 3.79 0.05 |
| BIDMC | 2.54 0.07 | 2.56 0.09 | 2.57 0.06 | 1.22 0.05 | 1.15 0.04 |
| MIMIC-III | 2.25 0.02 | 2.25 0.02 | 2.24 0.02 | 2.20 0.02 | 2.21 0.02 |
| Skin Temperature Estimation (MAE) | |||||
| Task/Dataset (MAE) | GPT-PPG | MOMENT-PPG | ||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-7 | 19M | 85M | 345M | 40M | 125M | 385M |
| SpO 2 Estimation (MAE) | ||||||
| Sleep Disorder | 3.82 0.08 | 3.88 0.08 | 3.90 0.08 | 3.27 0.07 | 3.28 0.07 | 3.20 0.06 |
| BIDMC | 1.48 0.05 | 1.04 0.04 | 1.18 0.04 | 1.62 0.05 | 1.46 0.05 | 1.15 0.04 |
| MIMIC-III | 2.46 0.03 | 2.39 0.03 | 2.43 0.03 | 2.20 0.03 | 2.20 0.03 | 2.20 0.03 |
| Skin Temperature Estimation (MAE) | ||||||
| Task/Dataset (MAE) | MOMENT | Chronos-2 | |||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-6 | 40M | 125M | 385M | 28M | 120M |
| Potassium Estimation (MAE) | |||||
| MIMIC-III | 0.47 0.004 | 0.47 0.004 | 0.47 0.004 | 0.47 0.004 | 0.47 0.004 |
| Inst. A | 0.45 0.01 | 0.45 0.01 | 0.45 0.01 | 0.47 0.01 | 0.52 0.01 |
| Sodium Estimation (MAE) | |||||
| MIMIC-III | 3.63 0.03 | 3.63 0.03 | 3.63 0.03 | 3.66 0.03 | 3.67 0.03 |
| Task/Dataset (MAE) | GPT-PPG | MOMENT-PPG | ||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-7 | 19M | 85M | 345M | 40M | 125M | 385M |
| Potassium Estimation (MAE) | ||||||
| MIMIC-III | 0.47 0.00 | 0.46 0.00 | 0.49 0.00 | 0.46 0.00 | 0.46 0.00 | 0.46 0.00 |
| Inst. A | 0.49 0.01 | 0.48 0.01 | 0.53 0.01 | 0.47 0.01 | 0.58 0.01 | 0.64 0.01 |
| Sodium Estimation (MAE) | ||||||
| MIMIC-III | 4.08 0.03 | 4.08 0.04 | 3.98 0.04 | 3.75 0.04 | 3.75 0.03 | 3.74 0.04 |
| Score | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG | |||||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-6 \ltx@saved@cline 7-9 \ltx@saved@cline 10-12 | 40M | 125M | 385M | 28M | 120M | 19M | 85M | 345M | 40M | 125M | 385M |
| Full suite — classification: head-only tuning | |||||||||||
| Family score | 4/6 | 0/6 | 2/6 | 0/6 | |||||||
| Win score | 0/6 | 1/6 | 3/6 | 0/6 | 0/6 | 0/6 | 2/6 | 0/6 | 0/6 | 0/6 | 0/6 |
| Combined | Generalists: 4/6 | Specialists: 2/6 | |||||||||
| Full suite — classification: full-model tuning | |||||||||||
| Generalists | Specialists | |||
| \ltx@saved@cline 2-3 \ltx@saved@cline 4-5 Actual task | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG |
| Head / Full | Head / Full | Head / Full | Head / Full | |
| Classification: F1 or accuracy (%), | ||||
| AF detection | 59.95 / 73.86 | 61.77 / 76.80 | 68.14 / 76.15 | 65.45 / 74.54 |
| Mental workload | 69.60 / 69.77 | 66.43 / 68.96 | 62.94 / 64.38 | 60.84 / 61.17 |
| Activity recognition | 78.37 / 86.39 | 81.33 / 77.14 | 66.14 / 78.09 | 70.61 / 72.09 |
| Generalists | Specialists | |||
| \ltx@saved@cline 2-3 \ltx@saved@cline 4-5 Actual task | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG |
| Head / Full | Head / Full | Head / Full | Head / Full | |
| Classification: 4 task categories; F1 or accuracy (%), | ||||
| AF detection | 59.95 / 73.86 | 61.77 / 76.80 | 68.14 / 76.15 | 65.45 / 74.54 |
| Mental workload / stress | 69.60 / 69.77 | 66.43 / 68.96 | 62.94 / 64.38 | 60.84 / 61.17 |
| Activity recognition | 78.37 / 86.39 | 81.33 / 77.14 | 66.14 / 78.09 | 70.61 / 72.09 |
| Specialist pretraining- source overlap | Downstream scale | Train hours | Specialist advantage | ||
| \ltx@saved@cline 5-6 | Small tier | Medium tier | |||
| Inst.A | All | 13 | 105.0 [99.0–381.0] | +1.2 [ 4.5–4.1] | 2.1 [ 9.3–5.8] |
| Non-Inst.A | h | 14 | 3.8 [0.2–9.8] | +3.2 [ 3.9–25.4] | +0.1 [ 6.8–2.7] |
| Non-Inst.A | – h | 5 | 23.0 [12.6–94.2] | +16.5 [1.3–17.6] | +1.2 [1.2–32.6] |
| Non-Inst.A | h | 13 | 100.0 [100.0–322.3] | +1.1 [ 2.0–8.0] | +1.0 [ 1.2–2.5] |
| Objective | Transformer | Convolution | State-space | LRU |
| Masked | ||||
| Next-patch | ||||
| Contrastive |
| Dataset | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG |
| Stanford | 68.20 1.28 [65.61, 70.61] | 68.57 1.27 [66.09, 70.96] | 81.25 1.07 [79.13, 83.31] | 64.69 1.27 [62.15, 67.21] |
| Simband | 54.17 4.15 [46.07, 62.69] | 65.81 3.95 [57.71, 73.51] | 50.00 4.42 [40.94, 58.59] | 47.62 4.47 [38.67, 55.88] |
| MAUS | 55.39 2.80 [49.76, 60.91] | 57.92 2.80 [52.54, 63.31] | 64.86 2.48 [59.78, 69.62] | 61.34 2.64 [55.84, 66.37] |
| Mental Stress | 44.61 8.23 [28.07, 59.70] | 59.20 7.41 [44.44, 73.02] | 61.76 7.16 [46.87, 74.67] | 73.53 6.14 [59.65, 83.96] |
| Pulse Transit | 82.69 0.97 [80.75, 84.66] | 78.13 1.10 [76.05, 80.34] | 73.10 1.34 [70.57, 75.72] | 63.50 1.39 [60.74, 66.16] |
| Real-world | 73.07 1.68 [69.85, 76.29] | 25.00 1.67 [21.71, 28.29] | 76.57 1.65 [73.29, 79.72] | 65.71 1.79 [62.29, 69.29] |
| Task | Dataset | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG |
| SpO 2 Estimation | MIMIC-III | 2.41 0.03 [2.36, 2.46] | 2.44 0.02 [2.39, 2.48] | 2.44 0.03 [2.40, 2.50] | 2.41 0.03 [2.36, 2.46] |
| Heart Rate Estimation | WESAD | 10.95 0.08 [10.80, 11.11] | 11.01 0.08 [10.85, 11.16] | 11.43 0.08 [11.27, 11.60] | 11.21 0.08 [11.05, 11.37] |
| Heart Rate Estimation | DALIA | 12.27 0.11 [12.07, 12.50] | 14.66 0.12 [14.44, 14.90] | 18.44 0.14 [18.18, 18.70] | 12.67 0.11 [12.47, 12.88] |
| Respiration Rate Estimation | MIMIC-III | 5.03 0.04 [4.95, 5.11] | 5.02 0.04 [4.94, 5.10] | 5.05 0.04 [4.97, 5.14] | 5.00 0.04 [4.92, 5.08] |
| Skin Temperature Estimation | MIMIC-III | 1.89 0.02 [1.86, 1.92] | 1.87 0.01 [1.84, 1.89] | 1.72 0.02 [1.69, 1.75] | 1.91 0.02 [1.87, 1.94] |
| Diastolic BP Estimation | PPG-BP | 10.00 0.66 [8.78, 11.38] | 9.25 0.68 [7.98, 10.66] | 9.18 0.69 [7.89, 10.60] | 9.68 0.72 [8.32, 11.09] |
| Task (bline type) | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG | Statistical baseline |
| Blood pressure estimation (Diastolic/Systolic) (MAE) (Age,Height,Weight) | |||||
| BUT-PPG | 8.46 0.13 / 16.61 0.35 | 7.51 0.14 / 14.01 0.28 | 7.71 0.16 / 15.43 0.34 | 7.70 0.14 / 14.50 0.30 | 6.24 0.13 / 7.29 0.19 |
| Heart rate estimation (MAE) (Age,Height,Weight) | |||||
| BUT-PPG | 10.36 0.30 | 10.05 0.29 | 10.22 0.28 | 10.27 0.28 | 10.67 0.31 |
| Troponin estimation (MAE) (Carry forward) | |||||
| MIMIC | 1.11 0.02 | 1.26 0.02 | 1.17 0.02 | 1.23 0.02 | 0.42 0.04 |
| Task / Subset | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG | Carry forward |
| Troponin Estimation (MAE) | |||||
| all | |||||
| pos | |||||
| neg | |||||
| pos2neg | |||||
| neg2pos | |||||
| Task/Dataset (Metric) | MOMENT | Chronos-2 | GPT-PPG (w/ & w/o Laplace) | MOMENT-PPG | |
| Train set | Test set | ||||
| AF Detection (F1 score) | |||||
| Stanford | Simband | ( ) | |||
| Simband | Stanford | ( ) | |||
| SpO 2 Estimation (MAE) | |||||
| Sleep Disorder | MIMIC | ( ) | |||
| Task/Signal (Metric) | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG |
| ICP Prediction (MAE) | ||||
| TCD CBFV signal | ||||
| Blood Pressure Estimation (Diastolic/Systolic MAE) | ||||
| ECG signal | / | / | / | / |
| Sleep Detection (F1 score) | ||||
| Encoder | Regression (15), | Solvable (9) | Classif. (6), F1 |
| Generalist checkpoints | |||
| MOMENT-Small | [ , ] | ||
| MOMENT-Base | [ , ] | ||
| MOMENT-Small, stat.-matched | [ , ] | ||
| Chronos-2 | [ , ] | ||
| After PPG pre-training | |||
| Step 0 | 2,000 | 6,000 | 10,000 | End | |
| MOMENT, tasks better (of 15) | |||||
| MOMENT, median gain (%) | |||||
| Chronos-2, tasks better (of 15) | 14 | 12 | 9 | 9 | 6 |
| Chronos-2, median gain (%) |
| MOMENT | Chronos-2 | |||||
| \ltx@saved@cline 2-4 \ltx@saved@cline 5-7 Step | Off. | Rand. | Part. | Off. | Rand. | Part. |
| 2,000 | ||||||
| 4,000 | – | – | – | |||
| 6,000 | ||||||
| 10,000 | ||||||
| 16,000 | – | – | – | |||
| Panel A: Supervised heart-rate estimation under controlled female:male training-window ratios | |||||
| Train source | Balanced test | MOMENT | Chronos-2 | GPT-PPG | MOMENT-PPG |
| MIMIC | MIMIC | [3pt] : | [3pt] : | [3pt] : | [3pt] : |
| MIMIC | BUT-PPG | [3pt] : | [3pt] : | [3pt] : | [3pt] : |
| MIMIC | DALIA | [3pt] : | [3pt] : | [3pt] : | [3pt] : |
| BUT-PPG | BUT-PPG | [3pt] : | [3pt] : | [3pt] : | [3pt] : |
| BUT-PPG | MIMIC | [3pt] : | [3pt] : | [3pt] : | [3pt] : |
Time-Series Foundation Models
Time-series foundation models are pretrained sequence models designed to transfer across forecasting, classification, anomaly detection, and other temporal tasks.
See all 343 papers