Who leads on each benchmark

Per-window 10 s, deterministic RR-variability tools, no training; sens 78.3 / spec 56.1; 44 pts; only row on this axis

External single-lead AF

Same pipeline on ECG; sens 71.9 / spec 63.6; 25 pts; only row on this axis

Window-level six-class Macro AUROC (mean ± 0.0011, 3 seeds) on public Liu PPG arrhythmia dataset; multi-class rhythm, not AF-vs-rest

ECG–PPG multi-class rhythm; SQI-anchored LLM (GPT-5.2); not AF-only

MIMIC 8 s AF, frozen features; accuracy only (no AUROC) — not comparable to AUROC rows; Chronos 0.81 / NormWear 0.78 / PaPaGei 0.66

PTB-XL Rhythm AFIB label, 12-lead, linear probe, 100% labels (Table 3 / D1); MIMIC-IV-ECG pretrain

DeepBeat wrist-PPG AF, full finetune, 5-run mean ±0.019 (npj 2026 BenchECG); same protocol ECGFounder 0.846, ECG-CPC 0.832

Frozen ECG-pretrained encoder on wrist PPG, 5-run mean ±0.013

MIMIC-BP SDE SBP/DBP; clinical/lab PPG

PPG-BP (Liang 2018) 2 s segments, frozen features (paper Table 1); finger clinical PPG; DBP Chronos 8.37 better

Smart-ring PPG, controlled CV perturbations; LOSO; ME 1.08/0.09

UCI cuff-less BP 942 subj; 10 s @ 125 Hz; lab PPG

UCI-BP transfer SDE; clinical/lab

CABiGRU pre-trained on diffusion-generated smartwatch IMU windows, fine-tuned on real data

Average across 6 subject-disjoint benches — not a single-benchmark rank

LOSO, right wrist 50 Hz 2.56 s, NULL kept; comparator Yuan et al. UKB ResNet18 (harnet-ukb) 77.06 LP / 84.00 FT on same protocol (paper Table 5).

LOSO, 60 participants, offline window classification, 6 microgestures + unknown; μDPad intermediate-fusion VGG-style 1D CNN (670k params), Acc+Gyro+PPG(green+IR); value_secondary = F1 77.0. Online HMM: 78.6% acc, 5.3% FPR (abstract states 3.7% FPR)

PPG-only single selected λ; participant-macro; routing off; ICLR'27 sub

Paper Table 1 frozen

8 s windows, frozen features (paper Table 1); split unstated; same-paper re-runs NormWear 10.16 / PaPaGei 12.25 / Chronos 12.34

Frozen linear-probe; paper protocol

PPG-only — not comparable to PPG+ACC SOTA rows above; ICLR'27 sub

Paper linear-probe

Under Pulse-PPG protocol (not PaPaGei original)

WESAD subject-independent 5-fold CV; mean Accuracy 90.34%; Stress Recall 94.07%

Binary stress, 8 s, frozen features (paper Table 2); accuracy — not comparable to AUROC rows; PaPaGei 0.80 / Chronos 0.77 / NormWear 0.77

WESAD subject-independent 5-fold CV; dual ECG+PPG encoders + cross-modal correlation fusion; paper AUC-ROC 97.29%

Binary stress; paper non-stress definition

LOSO pooled AUC, BVP 64 Hz 60 s windows, stress from 9 subjects; post hoc; interval-only 0.911–0.933; RF 0.972 transductive z-scoring (not ranked)

binary stress; classical ML on HRV features; PPG features linearly calibrated to ECG space (uncalibrated PPG 0.79; ECG 0.85); 15 subjects

Under Pulse-PPG protocol

model=AnyPPG; frozen encoder probe, 6,348 encounter-matched ED encounters; falls to 0.05 after age + renal adjustment

HMC sleep staging, EEG, frozen features + strict linear probe, mean ± sd over 3 pretrain seeds (±0.0045); full fine-tune 0.7669

ISRUC subgroup I, EEG, frozen features + strict linear probe, mean ± sd over 3 pretrain seeds (±0.0081); full fine-tune 0.8107