Deriving Health Metrics from the Photoplethysmogram: Benchmarks and Insights from MIMIC-III-Ext-PPG

Mohammad Moulaeifard; Philip J. Aston; Peter H. Charlton; Nils Strodthoff

cs.LGMar 23, 2026

Deriving Health Metrics from the Photoplethysmogram: Benchmarks and Insights from MIMIC-III-Ext-PPG

Authors: Mohammad Moulaeifard, Philip J. Aston, Peter H. Charlton, Nils Strodthoff

Organizations: AI4Health Department, Oldenburg University, Oldenburg, Germany · Department of Data Science and AI, National Physical Laboratory, Teddington, United Kingdom · School of Mathematics and Physics, University of Surrey, Guildford, United Kingdom · Department of Public Health and Primary Care, University of Cambridge, Cambridge, United Kingdom

Abstract

Photoplethysmography (PPG) is one of the most widely captured biosignals for clinical prediction tasks, yet PPG-based algorithms are typically trained on small-scale datasets of uncertain quality, which hinders meaningful algorithm comparisons. We present a comprehensive benchmark for PPG-based clinical prediction using the \dbname~dataset, establishing baselines across the full spectrum of clinically relevant applications: multi-class heart rhythm classification, and regression of physiological parameters including respiratory rate (RR), heart rate (HR), and blood pressure (BP). Most notably, we provide the first comprehensive assessment of PPG for general arrhythmia detection beyond atrial fibrillation (AF) and atrial flutter (AFLT), with performance stratified by BP, HR, and demographic subgroups. Using established deep learning architectures, we achieved strong performance for AF detection (AUROC = 0.96) and accurate physiological parameter estimation (RR MAE: 2.97 bpm; HR MAE: 1.13 bpm; SBP/DBP MAE: 16.13/8.70 mmHg). Cross-dataset validation demonstrates excellent generalizability for AF detection (AUROC = 0.97), while clinical subgroup analysis reveals marked performance differences across subgroups by BP, HR, and demographic strata. These variations appear to reflect population-specific waveform differences rather than systematic bias in model behavior. This framework establishes the first integrated benchmark for multi-task PPG-based clinical prediction, demonstrating that PPG signals can effectively support multiple simultaneous monitoring tasks and providing essential baselines for future algorithm development.

Figures & tables

Metric	Heart rhythm	RR	HR	BP
Number of subjects	6,189	4,695	5,924	2,391
Number of segments (30s)	6,399,754	5,546,169	5,187,183	2,468,693
Total Duration (h)	$\sim$ 53,331	$\sim$ 46,218	$\sim$ 43,227	$\sim$ 20,572
Age (years, mean $\pm$ SD)	64.1 $\pm$ 17.0	65.0 $\pm$ 16.1	63.9 $\pm$ 16.9	63.2 $\pm$ 16.2
Weight (kg, mean $\pm$ SD)	82.2 $\pm$ 22.6	82.0 $\pm$ 21.8	82.3 $\pm$ 22.6	83.2 $\pm$ 21.9
Height (cm, mean $\pm$ SD)	169.5 $\pm$ 10.5	169.7 $\pm$ 10.5	169.4 $\pm$ 10.5	169.7 $\pm$ 10.4

TABLE I : Demographic and clinical characteristics of the study population across different physiological subsets. Values are presented as mean

\pm

SD or percentage (%).

		AUROC		Sensitivity (Specificity $>$ 0.8)		Specificity (Sensitivity $>$ 0.8)
Task	Rhythm	XResNet1d101	LeNet1D	XResNet1d101	LeNet1D	XResNet1d101	LeNet1D
AF	AF	0.96	0.94	0.92	0.92	0.98	0.96
AF	AFLT	0.77	0.78	0.62	0.64	0.61	0.62
SAA	SINUS	0.92	0.90	0.89	0.86	0.86	0.84
SAA	AF+AFLT	0.95	0.94	0.93	0.92	0.98	0.96
ARRH	SR	0.93	0.93	0.93	0.92	0.89	0.89

TABLE II : In-distribution classification performance for robust AF classification (AF task), sinus versus atrial arrhythmia discrimination (SAA task), and general arrhythmia classification (ARRH task) in terms of label-dependent AUROC for the two models considered in this work.

	Baseline	XResNet1d101			LeNet1D
Parameter	MAE	MAE	Bias	95% LoA	MAE	Bias	95% LoA
SBP [mmHg]	19.06	16.13	-2.58	[-43.16, 38.00]	16.65	-1.24	[-43.00, 40.52]
DBP [mmHg]	10.09	8.70	0.79	[-21.34, 22.92]	9.39	2.53	[-20.38, 25.45]
HR [bpm]	14.37	1.13	-0.14	[-8.00, 7.72]	1.61	0.03	[-8.77, 8.84]
RR [bpm]	5.34	2.97	-0.17	[-10.07, 9.72]	4.53	0.31	[-11.89, 12.50]

TABLE III : Regression performance including MAE, Bland–Altman bias, and 95% limits of agreement (LoA) for BP, HR, and RR obtained with a median baseline predictor, XResNet1d101, and LeNet1D.

Fig. 1 : External validation of the rhythm classification performance (using XResNet1d101) demonstrating robust generalization of PPG-based rhythm classification, with AF detection showing exceptional cross-dataset transferability.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Rhythm	Description	Patients	Samples (30s)	Rhythm per Tasks (mapped)
				AF (Task 1)	SAA (Task 2)	ARRH (Task 3)
SR	Sinus rhythm	5,316	3,950,724	OTHER	SINUS	SR
STACH	Sinus tachycardia	2,838	1,192,025	OTHER	SINUS	STACH
AF	Atrial fibrillation	1,132	597,769	AF	AF+AFLT	AF
SBRAD	Sinus bradycardia	1,608	198,432	OTHER	SINUS	SBRAD
VPACE	Ventricular pacing	282	126,810	OTHER	OTHER	VPACE

Appendix

TABLE IV : Rhythm mapping after applying sample and patient threshold rules. OTHER is not considered as a separate class, but samples are retained with zero labels for all outputs to improve robustness to non-target rhythms.

BP Category	Systolic (SBP)	Operator	Diastolic (DBP)
Hypotension	$<90$ mmHg	OR	$<60$ mmHg
Normal	$90$ – $119$ mmHg	AND	$60$ – $79$ mmHg
Elevated	$120$ – $129$ mmHg	AND	$<80$ mmHg
Stage 1 Hypertension	$130$ – $139$ mmHg	OR	$80$ – $89$ mmHg
Stage 2 Hypertension	$\geq 140$ mmHg	OR	$\geq 90$ mmHg

Appendix

TABLE V : BP classification criteria ( [ 58 , 31 ] ). Note that when systolic and diastolic blood pressure values fall into different categories, the hypertension stage is assigned according to the higher category, in accordance with established clinical guideline conventions [ 58 ] .

HR Category	HR (bpm)
Bradycardia	$<60$
Normal	$60$ – $100$
Tachycardia	$>100$

Appendix

TABLE VI : HR classification criteria for adults [ 32 ] . Note: sinus tachycardia is defined as a sinus node discharge rate

>100

times per minute, i.e., HR

>100

bpm.

Category	BMI (kg/m 2 )
Underweight	$<18.5$
Normal	$18.5$ – $24.9$
Overweight	$25.0$ – $29.9$
Obese	$\geq 30$

Appendix

TABLE VII : BMI categories according to the WHO [ 33 ] .

Liu et al. Rhythm	AF Task Mapping	SAA Task Mapping	ARRH Task Mapping
0: Sinus rhythm	OTHER	SINUS	SR
1: Premature ventricular contraction	Excluded	Excluded	Excluded
2: Premature atrial contraction	Excluded	Excluded	Excluded
3: Ventricular tachycardia	OTHER	OTHER	VTACH
4: Supraventricular tachycardia	OTHER	OTHER	SVTACH
5: AF	AF	AF+AFLT	AF

Appendix

TABLE VIII : Rhythm mapping between Liu et al. dataset [ 5 ] and MIMIC-III-Ext-PPG classification tasks

Study	Patients ( n )	Data Duration ( $\mathrm{h}$ )	SBP MAE ( $\mathrm{mmHg}$ )	DBP MAE ( $\mathrm{mmHg}$ )	Data Source/Subset
Ours	2,391	20,572	$\sim 16\text{--}17$	$\sim 9\text{--}10$	MIMIC-III ( MIMIC-III-Ext-PPG )
Moulaeifard et al. [ 10 ]	1,474	2,357	$\sim 15\text{--}17$	$\sim 9\text{--}10$	MIMIC-III (PulseDB [ 59 ] )
Sanches et al. (MIMIC-BP) [ 49 ]	1,524	380	$\sim\textbf{13\text{--}14}$	$\sim 9\text{--}10$	MIMIC-III (Custom Subset)

Appendix

TABLE IX : Comparison of BP Estimation Performance

Task	Rhythm	Baseline $\uparrow$	BP category $\uparrow$					HR category $\uparrow$
Task	Rhythm		HPN	Normal	Elevated	Stage 1 HTN	Stage 2 HTN	Bradycardia	Normal	Tachycardia
AF	AF	0.96	0.95	0.99	0.99	0.99	0.97	0.88	0.96	0.97
AF	AFLT	0.77	0.80	0.77	0.78	0.80	0.84	0.74	0.76	0.86
SAA	SINUS	0.92	0.88	0.95	0.96	0.97	0.96	0.84	0.91	0.96
SAA	AF+AFLT	0.95	0.96	0.99	0.99	0.99	0.97	0.87	0.95	0.97
ARRH	SR	0.93	0.92	0.95	0.96	0.96	0.94	0.72	0.86	0.76

Appendix

TABLE X : AUROC performance of heart rhythm detection across BP and HR categories (XResNet1d101). In-distribution baselines are from Table II . Dashes indicate insufficient samples. HTN: hypertension; HPN: hypotension.

Task	Rhythm	Baseline $\uparrow$	BP category $\uparrow$					HR category $\uparrow$
Task	Rhythm		HPN	Normal	Elevated	Stage 1 HTN	Stage 2 HTN	Bradycardia	Normal	Tachycardia
AF	AF	0.92 / 0.98	0.92 / 0.98	0.99 / 0.99	0.99 / 0.99	0.99 / 0.99	0.94 / 0.99	0.79 / 0.78	0.93 / 0.99	0.96 / 0.97
AF	AFLT	0.62 / 0.61	0.70 / 0.73	0.61 / 0.65	0.54 / 0.70	0.63 / 0.72	0.78 / 0.79	0.51 / 0.50	0.62 / 0.59	0.73 / 0.76
SAA	SINUS	0.89 / 0.86	0.77 / 0.78	0.97 / 0.91	1.00 / 0.95	0.99 / 0.96	0.99 / 0.93	0.70 / 0.74	0.87 / 0.84	0.97 / 0.95
SAA	AF+AFLT	0.93 / 0.98	0.92 / 0.98	0.99 / 0.99	0.99 / 1.00	0.99 / 1.00	0.95 / 0.99	0.80 / 0.80	0.92 / 0.98	0.96 / 0.98
ARRH	SR	0.93 / 0.89	0.93 / 0.88	0.95 / 0.94	0.97 / 0.95	0.96 / 0.95	0.94 / 0.92	0.51 / 0.49	0.75 / 0.76	0.58 / 0.58

Appendix

TABLE XI : Sensitivity (at specificity

\geq

0.8) and specificity (at sensitivity

\geq

0.8) across BP and HR categories (XResNet1d101). Values are reported as sensitivity / specificity. Baseline results are from Table II . Dashes indicate insufficient samples.

Task	Rhythm	Baseline $\uparrow$	Gender $\uparrow$		BMI $\uparrow$			Ethnicity $\uparrow$
Task	Rhythm		Female	Male	Normal	Overweight	Obese	White	Black	Asian	Hispanic
AF	AF	0.96	0.95	0.96	0.92	0.92	0.98	0.95	0.96	0.98	0.95
AF	AFLT	0.77	0.84	0.75	0.72	0.75	0.75	0.74	0.93	–	–
SAA	SINUS	0.92	0.91	0.93	0.84	0.93	0.96	0.93	0.93	0.98	0.82
SAA	AF + AFLT	0.95	0.95	0.95	0.93	0.93	0.97	0.95	0.95	0.98	0.95
ARRH	SR	0.93	0.93	0.92	0.87	0.92	0.95	0.93	0.95	0.98	0.91

Appendix

TABLE XII : Performance (AUROC) across gender, BMI, and ethnicity categories (XResNet1D101). Baseline results are taken from Table II . Dashes indicate insufficient samples. The ”Underweight” BMI category was excluded due to the insufficient number of samples.

Task	Rhythm	Baseline $\uparrow$	Gender $\uparrow$		BMI $\uparrow$			Ethnicity $\uparrow$
Task	Rhythm		Female	Male	Normal	Overweight	Obese	White	Black	Asian	Hispanic
AF	AF	0.92 / 0.98	0.92 / 0.98	0.93 / 0.98	0.89 / 0.94	0.87 / 0.92	0.97 / 0.99	0.90 / 0.97	0.95 / 0.98	0.96 / 0.99	0.93
AF	AFLT	0.62 / 0.61	0.74 / 0.74	0.60 / 0.59	0.49 / 0.53	0.54 / 0.60	0.70 / 0.52	0.92 / 0.58	0.89 / 0.91	–	–
SAA	SINUS	0.89 / 0.86	0.86 / 0.84	0.91 / 0.88	0.69 / 0.72	0.92 / 0.88	0.97 / 0.94	0.93 / 0.89	0.93 / 0.91	0.99 / 0.98	0.66 / 0.62
SAA	AF+AFLT	0.93 / 0.98	0.93 / 0.98	0.93 / 0.98	0.90 / 0.94	0.88 / 0.93	0.96 / 0.99	0.92 / 0.97	0.93 / 0.96	0.97 / 0.99	0.91 / 0.91
ARRH	SR	0.93 / 0.89	0.93 / 0.90	0.92 / 0.89	0.80 / 0.80	0.92 / 0.88	0.96 / 0.94	0.93 / 0.90	0.96 / 0.93	0.99 / 0.98	0.88 / 0.84

Appendix

TABLE XIII : Sensitivity (at specificity

\geq

0.8 ) and specificity (at sensitivity

\geq

0.8) across demographic groups (XResNet1d101). Values are reported as sensitivity / specificity. Baseline results are from Table II . Dashes indicate insufficient samples.

Original Ethnicity Label	Aggregated Group
WHITE	White
WHITE - RUSSIAN	White
WHITE - EASTERN EUROPEAN	White
WHITE - OTHER EUROPEAN	White
WHITE - BRAZILIAN	White
PORTUGUESE	White

Appendix

TABLE XIV : Mapping of original MIMIC-III ethnicity labels to aggregated ethnicity groups, based on U.S. Office of Management and Budget Statistical Policy Directive No. 15 (1997) [ 34 ] . It is worth noting that due to extremely small numbers of subjects in certain categories (e.g., Caribbean Island,

n=1

; American Indian/Alaska Native,

n=2

; Native Hawaiian/Pacific Islander,

n=2

; multiracial,

n=17

), these groups were consolidated into the “OTHER” category to ensure statistical stability of subgroup analyses.

Explore similar work

PPG-Bench: A Unified Benchmark for PPG Foundation Models
Oct 9, 2026Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra +2Photoplethysmography Time-Series Foundation Models
Mannas AI
A robust PPG foundation model using multimodal physiological supervision
Jun 5, 2026Eloy Geenjaar, Vince Calhoun, Scott Daly +4Multimodal Robustness Multimodal Pretraining
Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, USA · Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS), Georgia State University, Georgia Institute of Technology, Emory University, Atlanta, USA · Dolby Laboratories, San Francisco, USA
Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines
Jul 26, 2026Bo Wu, Haoling Wang, Zhuodiao Kuang +1Photoplethysmography Wearable Health Monitoring
Department of Informatics and Networked Systems University of Pittsburgh Pittsburgh, USA · Department of Biostatistics and Health Data Science University of Pittsburgh Pittsburgh, USA · Language Technologies Institute Carnegie Mellon University Pittsburgh, USA

Oct 9, 2026J/K move · Enter open · S save

PPG-Bench: A Unified Benchmark for PPG Foundation Models

Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra, Sandeep Singh+1

Mannas AI

Photoplethysmography Time-Series Foundation Models

Open paper AskPDF arXiv

Photoplethysmography (PPG) is increasingly recorded by wearable and clinical devices at enormous scale, and a growing number of PPG foundation models (FMs) have been developed to compute reusable representations for various downstream tasks. These models are difficult to compare: different evaluations use different datasets, task definitions, preprocessing, downstream heads, hyperparameter tuning, and baselines. We present PPG-Bench, a unifying open-source benchmark evaluating five public PPG FMs across nine open-access datasets and 26 tasks spanning cardiovascular state estimation, activity, sleep, screening, and identification. We evaluate models under four downstream regimes of increasing adaptability (linear probing, random forests, frozen MLP probes, and end-to-end finetuning), with model-specific hyperparameter optimization and subject-disjoint cross-validation. We also evaluate against three controls: random/constant baseline, a deterministic pulse-rate-variability (PRV) feature set, and a generic time-series FM (MOMENT). Across all 68 task-regime combinations, 35.3% show no meaningful FM advantage over the random/constant baseline, and in 83.8% no PPG-specific FM outperforms both PRV and MOMENT. We also find that added downstream head capacity generally improves performance, but can cause significant degradation if the model is performing at chance/constant level. Current PPG FMs therefore cannot yet claim to provide generalised and transferrable representations of PPG signals, and future evaluation in this field should include checks against simple, non-learned controls. Code for the benchmark and instructions to evaluate more models can be found here - https://github.com/mannas-ai/ppg-bench.

Abstract

Explore similar work

PPG-Bench: A Unified Benchmark for PPG Foundation Models

A robust PPG foundation model using multimodal physiological supervision

Blood Pressure Estimation from PPG: A Comparative Study of Direct and ECG-Mediated Deep Learning Pipelines