Estimating blood pressure from the electrocardiogram: findings of a large-scale negative results study

2.1. Database

The data used in this study consists of ECG and BP measurements from two databases collected over two years from August 2019 to March 2021 by AliveCor (Mountain View, CA, USA), using the following devices:

• (i)

  OMRON Complete (Omron Healthcare, Kyoto, Japan), which is an integrated BP monitor and single-lead ECG;

• (ii)

  KardiaMobile (AliveCor, Mountain View, CA, USA) for collecting single-lead ECGs and independent BP readings from portable BP devices (Omron Healthcare, Kyoto, Japan).

To note, the ECG and BP were self-recorded asynchronously in non-clinical settings, with variable numbers of BP and ECG per subject and varying time gaps between the two modalities (varying between seconds and hours). The ECG dataset comprises 180 790 records from 10 624 subjects, with a minimum time gap of 30 seconds between two consecutive ECG recordings for each unique subject. ECGs were recorded at a sampling frequency of 300 Hz. The BP dataset consists of 21 227 729 measurements, corresponding to 297 965 subjects. A total of 10 346 subjects, which were common between the ECG and BP datasets, were shortlisted for this study.

2.2. Data cleaning

The data cleaning process is summarized in figure 1. Accordingly, records were selected from the matched dataset based on the following criteria:

• (i)

  The analysis was limited to adult male and female subjects aged between 18 and 90 years at the time of ECG recording. Subjects with unknown sex or with age outside $[18,90]$ were excluded from the analysis.

• (ii)

  Records with misreported DBP values higher than SBP were removed. Then, thresholds were applied to define valid BP ranges. Valid BP ranges were set to DBP between 20–200 mmHg and SBP between 30–300 mmHg. These thresholds are consistent with the pre-processing approach used in our previous study, which analyzed approximately 75 million BP values from the general population (Mousavi et al 2024).

• (iii)

  The dataset included ECG classification labels generated by AliveCor’s proprietary ECG analysis software, which labels signals as ‘sinus rhythm’, ‘atrial fibrillation’, ‘bradycardia’, ‘tachycardia’, ‘unclassified’, ‘too short’, ‘unreadable’ and missing values. Records labeled as ‘unreadable’ or with missing labels were removed.

• (iv)

  For consistency, ECG record lengths were fixed to 30 s, and the records shorter than 30 s were excluded from the analysis. Previous studies indicate that this duration is sufficient for capturing essential ECG features, especially for rhythm analysis and heart rate variability (HRV) (Munoz et al 2015). 98% of the ECG database complied with this requirement. For consistency, the ECGs longer than 30 s were truncated to the first 30 s.

• (v)

  The ECG and BP data were collected asynchronously, resulting in varying time gaps between the ECG and BP measurements of the same subject. Given that both signals naturally fluctuate over time, we defined a maximum allowable time gap such that BP variability within this window would be minimal—ensuring that estimating BP from ECG remained both meaningful and clinically relevant. To determine this threshold objectively, we referred to acceptable BP error margins from BP device standards and reported rates of short-term BP variability in the literature. Presumably, as long as the time-gap between ECG and BP collection is within these thresholds, any BP change during that interval would fall within an acceptable error margin—making the ECG-BP pairing valid for estimation purposes.

  According to the Association for the Advancement of Medical Instrumentation (AAMI) standard, the mean BP error in BP measurement devices should be less than 5 mmHg (Stergiou et al 2018). To identify the time window during which a 5 mmHg change in BP might occur, relevant literature was reviewed. Most studies reported mean BP differences over 30 min or longer intervals (Mancia et al 1983, Graham et al 1995, Clement et al 2003, Kario et al 2003, Okamoto et al 2009, Sayk et al 2010, Mancia 2012). From these studies, reported mean BP differences and their corresponding time windows were extracted to estimate the ‘rate of BP variation’ over time. Using these rates, the time intervals corresponding to the negligible 5 mmHg change were calculated by dividing 5 mmHg by the rate of change. The resulting estimates ranged from 3.5 to 38 min. The minimum value (3.5 min), was considered as the acceptable short-term window, which we considered as the maximum allowable time gap between ECG and BP recordings.

• (vi)

  Many subjects had multiple BP measurements within the acceptable BP-ECG time interval window. For each subject and ECG, all BP measurements within the acceptable time window of (3.5 min) were averaged. Averaging BP measurements within short time windows is a standard procedure in clinical practice, which results in more accurate BP measurements (Mousavi et al 2024), and reduction of measurement biases (Nateghi and Sameni 2025).

The final filtered dataset included 124 427 pairs of BP and ECG records from 7412 subjects. Table 2 summarizes the distribution of the final dataset by sex and ECG labels, where ‘Normal’ represents ‘sinus rhythm,’ while all other labels including ‘atrial fibrillation’, ‘bradycardia’, ‘tachycardia’, and ‘unclassified’ are considered ‘Abnormal’. Also, table 3 presents the statistical distribution of BP datasets by sex and ECG labels, and figure 2 illustrates the 95% percentile range contours for males and females BP distribution. The mean SBP and DBP for each group are marked with dots.

2.3. ECG pre-processing

The ECG records were band-pass filtered with a band-pass frequency between 0.1 Hz and 100 Hz, and a notch filter at 50 or 60 Hz, depending on the local power line frequency. The notch filter was designed using a second-order infinite impulse response filter (iirnotch in MATLAB) with a quality factor (Q) of 40 and was applied.

2.4. Feature extraction

A total of 280 features were extracted, comprising 128 interpretable features extracted from the ECG records using a codebase developed by our team (Sameni 2006–2025); 150 features extracted using the Black Swan codebase (Zabihi et al 2017); the time gap between the ECG and the average time of the corresponding BP measurements (within 3.5 min time windows); and the subject age. To enable the replication of the implemented process, the complete feature set is described below.

• (i)

  Beat signal-to-noise ratio (SNR): To quantify beat-to-beat morphological consistency in the ECG over the 30 s segment, a SNR index was computed and assigned to each beat. R-peaks were first detected using the OSET robust R-peak detector function peak_det_likelihood (Sameni 2006–2025), and individual beats were segmented using a window of W samples centered around each R-peak. Robust weighted average (RWA) and robust beat median (RBM) beats were then calculated, following the method in Leski (2002). For each beat, the residual was computed as the difference from the RWA or RBM beat, and the beat SNR was defined as the power ratio between the original beat and the mean/median-based residuals. These SNRs capture both morphological deviations and measurement noise.

• (ii)

  HRV and HR metrics: After ECG R-peak detection, R-R intervals were computed and converted to instantaneous HR values in beats per minute (bpm). The HR sequence was next summarized using the mean, median, 5th percentile, and 95th percentile. HRV was assessed using the standard deviation of R-R intervals and the root mean square of successive differences (Clifford et al 2006).

• (iii)

  Time interval measurements: Fiducial points for each beat were extracted using the fiducial_det_lsim function from OSET (Sameni 2006–2025). Using these points, key ECG time intervals were calculated, including the QRS complex duration, QT interval, PR interval, ST interval, PR segment level, and ST segment level. Additional intervals were computed between specific peak pairs: P–R, Q–R, S–R, and T–R, to capture more detailed temporal relationships between waveform components. Corrected QT intervals (QTc) were also derived using the Bazett (QTc-B) and Fridericia (QTc-F) corrections (Luo et al 2004).

• (iv)

  Amplitude and morphological area metrics: Amplitude and area-based features were computed using fiducial points marking the onset, peak, and offset of each ECG waveform component. For each component, the amplitude and the area under the curve (sum of ECG values from onset to offset) were calculated. In addition, we computed the amplitude ratio of the R peak to other major peaks (P, Q, S, and T), and the amplitude difference between the S and T peaks across the ST segment.

• (v)

  Amplitude-to-timing ratios: For each beat, the difference between the R-peak amplitude and the amplitude of other peaks was divided by the time interval between the R peak and the corresponding peak, providing a measure of waveform shape (slope).

• (vi)

  Signal mobility and complexity: Mobility was computed as the ratio of the variance of the first derivative of the ECG to the variance of the ECG (Simjanoska et al 2018, 2020, Fuadah and Lim 2022). Complexity was calculated as the ratio of the variance of the second derivative to the variance of the first derivative, divided by the mobility value (Simjanoska et al 2018, 2020, Fuadah and Lim 2022).

• (vii)

  Singular value decomposition (SVD) metrics: SVD has been shown to encode ECG beat variability (Zheng et al 2021). ECG beats were segmented around each R-peak with a window of the median beat-to-beat interval, and stacked to form a 2D matrix (number of beats times number of samples of the segmented beats) using the event_stacker function from OSET (Sameni 2006–2025), where each row represents one beat. SVD was then applied to this matrix to extract singular values. The resulting values were normalized by the largest singular value and used as features to capture the similarity and reproducibility of ECG beats across the segment. The number of non-zero singular values of a rectangular matrix is smaller than or equal to the minimum of its rows and columns, which in our case was the number of beats used to construct the stacked beat matrix. To ensure a fixed feature length across all subjects and records, the SVD-based feature vector was set to a length of 45, corresponding to the maximum number of beats over 30 s across all subjects. Shorter vectors were zero-padded to reach this length.

• (viii)

  Black-Swan: This set includes 150 features developed by a top-performing team in the PhysioNet Challenge 2017 for atrial fibrillation classification (Zabihi et al 2017). The features span multiple domains, including time, frequency, time-frequency, phase space, and meta-level representations. This set has also been successfully applied in other ECG classification tasks (Bahrami Rad et al 2021, 2024, Koscova et al 2024).

The amplitude, interval, and morphological features described above were computed per beat. These beat-wise values were then summarized using the mean, median, and SD to form fixed-length feature vectors.

2.5. ML models

Decision tree-based regression models were used for their performance, their ability to handle feature sets with missing values, and their capacity to model complex and nonlinear relationships in data (Podgorelec et al 2002). This includes extreme gradient boosting, random forest (RF), CatBoost, and light gradient boosting machine (LightGBM).

2.6. Model developing and data splitting

Our previous studies have shown that, at the population level, males exhibit higher BP than females (Mousavi et al 2024). Therefore, our SBP and DBP estimation models were trained separately for each sex group. Furthermore, for each sex group, two distinct BP models were trained using: (1) only normal-labeled ECG records and (2) all records. This allowed us to investigate whether BP estimation performance differs when trained exclusively on normal ECGs versus both normal and abnormal cases. As a result, four distinct models were developed for each of SBP and DBP (male-normal/all and female-normal/all). See table 3 for the breakdown.

For training and validation, we used subject-level data splitting rather than record-level to avoid inter-subject data leakage between training and validation, ensuring that the results are generalizable to other datasets. Accordingly, all models were trained using leave-subject-out five-fold cross-validation. The preprocessed dataset (detailed in figure 1) was randomly split into five sets of subjects. In each fold, the model was trained on data from four sets and tested on the left-out set. In this cross validation scheme, each subject appeared only once in the test set and four times in the training set (i.e. in 4/5 folds). In terms of ECG-BP recording pairs, the subjects had varied numbers of recordings: over 26% of the subjects had only one pair; the median was six pairs per subject; and, in an extreme case, one subject had 3689 measurement pairs. To address this imbalance and reduce the risk of biasing the ML models toward subjects with more measurements, the number of ECG-BP pairs per subject was capped at six in the train and test datasets. For subjects with more than six recordings, six pairs were randomly selected during cross-validation to make the best use of the available data. For subjects having six or fewer ECG-BP pairs, the same pairs were used for training across all folds. Therefore, most of the 124 427 pairs of ECG-BP measurements listed in figure 1 eventually contributed to the training/testing procedure.

In each fold, the model predictions were stored and after completing all five folds for each model, the predictions from all folds were aggregated to calculate and report the performance metrics. The training and test sets used for training each fold were identical across all the studied regression/classification models to accomplish a fair comparison.

2.7. Evaluation metrics

The performance of the developed ML models was evaluated using various metrics, including mean error (ME), SD of ME, MAE, SD of MAE, and correlation coefficient, to enable comparison with other studies. Specifically, the correlation coefficient reflects the strength of the linear relationship between the estimated and actual BP values. The correlation coefficient can be either positive or negative, implying a direct or inverse relationship. The absolute correlation coefficient ranges from 0, indicating no linear relationship, to 1, indicating a perfect relationship (Martin Bland and Altman 1986).

3.1. ECG-based BP estimation

Table 4 summarizes the performance of the ML models in estimating DBP and SBP using only ECGs, based on sex and ECG labels (across all and normal-only ECG records). The best results, based on the correlation coefficient metric, were achieved in estimating DBP with a value of 0.38 using CatBoost and normal-ECG records of males, and in estimating SBP with a value of 0.35 using RF and all-ECG records of females.

Figure 3 illustrates the results of the prediction errors distribution (the difference between predicted and actual BP values) and 95% percentile contour plots of predicted vs. actual BP for the best-performing estimation models. In an unbiased and well-performing model, the predicted BP values should closely match the actual value, and the prediction errors should have a mean of zero—ideally exhibiting a symmetric unbiased distribution around this mean. However, in figure 3, we can see a non-zero mean and skewed error distribution, indicating a systematic bias and asymmetric error.

The performance of the proposed models was assessed using two widely recognized BP evaluation standards: the AAMI and the British Hypertension Society (BHS) (Stergiou et al 2018). According to the AAMI standard, a valid BP measurement model must achieve a ME of $\text{⩽}$5 mmHg and a SD of ME $\text{⩽}$8 mmHg. The BHS standard, in contrast, grades BP measurement devices based on the cumulative percentage of predictions within 5, 10, and 15 mmHg, assigning Grades A, B, or C accordingly. Based on these criteria, neither the SBP nor the DBP prediction models we developed on our dataset satisfied the AAMI requirements, as both exhibited ME and SD values exceeding the thresholds. With respect to the BHS grading, the SBP model failed to meet the standard, while the DBP model fulfills Grade C performance (table 5).

3.2. Subject-wise vs. record-wise model training

To evaluate the effect of data partitioning strategies on model performance, we conducted an additional experiment using record-wise cross-validation—where training and validation data were randomly split across individual records, regardless of subject identity (table 6). In this setting, the models achieved significantly higher correlation coefficients of 0.59 for SBP and 0.63 for DBP using RF, compared to 0.29 to 0.37 in the subject-wise setup described earlier. This increase in performance suggests that when data splitting is not performed correctly (i.e. using record-wise instead of subject-wise splitting), the models may be leveraging subject-specific patterns seen during training, rather than learning generalizable physiological relationships between ECG and BP that would transfer to unseen subjects.

3.3. ECG-based BP classification

We further investigated the relationship between ECG and BP within a classification framework, which generally, depending on the dataset, may be an easier or more difficult task than continuous BP value estimation (see (Hastie 2009), chapters 2 and 3)⁴. According to the American Heart Association, adult BP is diagnostically categorized into four groups: Normal, Elevated, Stage 1 Hypertension, and Stage 2 Hypertension (Reboussin et al 2018). Table 7 lists these categories along with the corresponding SBP and DBP ranges for each group.

Similar to the continuous SBP/DBP estimation scenarios, four decision tree-based classification models were trained, this time aiming to predict BP categories rather than the continuous BP values. To obtain a relatively balanced dataset, we considered a binary classification task: Normal and Elevated BP were grouped as the Normal (more specifically, non-hypertensive) class, and Hypertension Stages 1 and 2 were grouped as the Hypertensive class. Separate models were trained for males and females. With this grouping, the ratio of Hypertensive to Normal was 62% to 38% in males and 57% to 42% in females. The models were developed using the same 280 feature set as the regression problem (i.e. 278 ECG-based features, the average time gap between each ECG and its corresponding BP measurement, and the subject’s age). The classifiers were trained and validated using a subject-wise approach, and the number of records per subject was limited to a maximum of six (the median of the number of ECG-BP pairs across the dataset). The best models for males and females was CatBoost. The five-fold cross-validation, based on metrics including accuracy, F1 -score, specificity, positive predictive value (PPV), and negative predictive value (NPV), area under the receiver operating characteristic (AUROC) curve and area under the precision-recall (AUPR) curve showed minimal variation across the folds, demonstrating the consistency and robustness of the results.

Table 8 summarizes the results of the developed models by sex and at a sensitivity of 0.7. The best models, Catboost achieved AUROC of 0.636 and 0.655 for male and female datasets, respectively. Figure 4 shows the ROC and PR curves of the models for both groups. The ROC curves of the models exhibit very similar patterns across most sensitivity points. Differences in the PR curves are related to the differences in the ratios of Normal and Hypertensive classes in the male and female datasets (Sameni 2025). Overall, the classification models show poor performance in prediction of BP groups (Normal versus Hypertensive).

4.1. Model performance and generalization

Model performance on the validation set was low, with correlation coefficients around 0.30, indicating poor generalizability.

Apparently, the models consistently predicted BP values centered around the dataset mean. To explore this, we compared the model outputs with fixed values derived from the training set’s mean, median, and mode. As shown in table 9, the ML model predictions were nearly indistinguishable from simply using the median BP value, indicating that the models were effectively regressing to the mean. This phenomenon, known as central tendency bias or regression to the mean (Barnett 2004), occurs when a model lacks informative input features. In this case, the model appears to ignore ECG variability and rely instead on the statistical distribution of BP in the training data. Formally, this behavior suggests:

where $f(\cdot )$ denotes the probability density function, indicating that the ECG features contribute little to the conditional BP distribution.

4.2. Comparison with prior studies and standards

Our findings support earlier studies, which debated that accurate ECG-based BP estimation is unfeasible (Sato 2021, Landry and Mukkamala 2023), and differ from several earlier studies that reported promising results for ECG-based BP estimation (Mousavi et al 2018, Simjanoska et al 2018, 2020, Fan et al 2020, 2021, Miao et al 2020, Banerjee et al 2022, Wuerich et al 2022, Aldein et al 2023, 2025, Syah et al 2023, Kuzmanov et al 2024), as summarized in table 1. Accordingly, many of these prior studies relied on small sample sizes, lacked subject-level separation between training and testing, or used limited feature sets. The current study, using a large ambulatory dataset and rigorous subject-level validation, offers a more generalizable and conservative evaluation.

To assess the potential for clinical utility, we evaluated our models using two widely accepted standards: the AAMI and BHS. None of the models satisfied the AAMI thresholds (ME $\text{⩽}$ 5 mmHg, SD $\text{⩽}$ 8 mmHg), and only the DBP models marginally achieved Grade C according to the BHS scale. These results further emphasize that ECG-only BP estimation does not meet the performance required for clinical use.

4.3. Importance of across-subject validation

The subject-wise versus record-wise partitioning of training and validation records is another aspect often undocumented or overlooked in prior studies. Our results show that record-wise validation can significantly inflate model performance by allowing data from the same subject to appear in both training and test sets. This causes subject-level information leakage, particularly when some individuals contribute many records. As a result, the model learns subject-specific patterns rather than general physiological relationships between ECG and BP, performing well on familiar data but failing to generalize to new subjects. In contrast, our subject-wise approach—where all data from each subject was isolated to either training or validation—prevented this leakage and revealed the true complexity of the task. We also limited the number of records per subject to reduce bias from overrepresented individuals. The lower correlation coefficients in this setup provide a more realistic picture of model performance and reflect the actual difficulty of ECG-based BP estimation.

4.4. Impact of biological sex on BP

On a population level, males generally have higher average BP values than females (Mousavi et al 2024). Our dataset aligns with this result: the mean SBP values were 127.1 mmHg for males and 123.8 mmHg for females, while the mean DBP values were 80.2 mmHg and 78.9 mmHg, respectively. However, the SD of BP values in females was higher than in males, which may be the result of the imbalance number of records between the two groups. Therefore, BP estimation models were developed separately for male and female groups to account for sex-specific physiological differences. Although overall model performance on the validation set was poor (table 4), the correlation coefficient for SBP prediction was higher in females than in males. However, for DBP, the correlation coefficients were very similar across both sex groups. This may be due to the narrower range of DBP values (40–165 mmHg) compared to SBP values (60–227 mmHg). Furthermore, the 95% percentile contour plots (figure 3), which compare predicted versus actual BP values based on sex and ECG labels, indicate that the models predicted a broader prediction range for female subjects, in SBP estimation. This may be due to greater variability or fewer number of records from females in our dataset.

4.5. Physiological and statistical implications

From a physiological standpoint, the results can be explained by the fact that BP reflects vascular compliance, peripheral resistance, blood volume, and autonomic tone-factors that are not directly encoded in the electrical activity captured by the ECG. Statistically, this aligns with the concept of parameter identifiability in regression problems (Sameni 2023). Even with a large dataset and highly expressive models, some outputs remain fundamentally non-identifiable from a given input modality. Our findings suggest that BP estimation from ECG alone may fall into this category.

4.6. Limitations and future work

Several limitations should be acknowledged. First, only single-lead ECGs were used, limiting the available morphological and spatial information. Second, ECG and BP were not recorded simultaneously, though the time gap was constrained to a 3.5 min window and included as a model input. Additionally, all BP values were obtained using non-invasive home devices, which can introduce measurement noise and inaccurate cuff placement and subject positioning. Finally, our dataset may not have fully captured extreme (very low and very high) BP values, which could limit the generalization of our regression and classification models to extreme cases.

Future work should investigate larger BP datasets with full measurement ranges to better capture extremes and assess their impact on model performance. Future work should investigate models that integrate ECG with additional synchronous physiological signals, such as PPG, impedance cardiography, or accelerometry. Multi-lead ECG recordings may also provide more discriminative features. Furthermore, instead of predicting instantaneous BP values, it may be more plausible to estimate average BP over longer time windows. Finally, DL models applied directly to ECG waveforms could be explored, although the identifiability limitation observed in this study may still persist.