Deep learning prediction of nocturnal hypertension for patients intolerant to ambulatory blood pressure monitoring.

Figures

• 

  Fig. 1. Entropy of SBP-HR data point bin assignments before and after entropy reduction preprocessing. a Line plot for the Daytime ABPM subset (ID: 1620#2021-03-20). The blue line represents SBP recorded between 6:00 AM and 10:00 PM, and the yellow line depicts HR over the same time. b Bar plot showing entropy of sample points in the ABPM subset. Blue bars: SBP; Yellow bars: HR. c Left: KDE-based probability density plot of SBP-HR; darker colors indicate higher density. Right: Corresponding entropy heatmap from Monte Carlo simulation using 10-bin method; darker colors indicate lower entropy. Green points show original data. d Left: Cropped KDE density plot excluding grid points with density below 5% of maximum. Right: Corresponding entropy heatmap. Green points show original SBP-HR data points.

• 

  Fig. 2. Activity and posture transition probabilities as low-dimensional embeddings show clustering by statistical metrics. a The left panel illustrates the Markov chain for the example sequence “110,010,” where black triangular arrows indicate temporal progression. When these transitions are rearranged into a cyclically repeating pattern, as shown in the right panel, the Markov chain reaches its steady-state distribution. b Once the two-state Markov chain reaches steady state, the transition probabilities P01 and P10 can be used to derive P11, P00, the proportions of “1''s (P1) and “0''s (P0), and their respective average run lengths (L1 and L0). In effect, the two transition probabilities encapsulate six statistical metrics. Here, P01 denotes the probability of transitioning to state 1 given that the current state is 0, and P10 denotes the probability of transitioning to state 0 given that the current state is 1. c The left panel presents the two-dimensional distribution of P01-P10 for all samples' posture time series. Each sample is assigned a three-digit binary state code based on the following criteria: the first digit is set to 1 if the sequence contains more than 50% “1''s; the…

• 

  Fig. 3. ABPM-VAE Architecture and Data Flow. The ABPM-VAE model consists of two interconnected pathways that jointly predict nocturnal BP-HR probability density maps from preprocessed ABPM data. Upper pathway (reconstruction branch): The input BP-HR probability density map is passed through multiple convolutional layers to extract features and generate latent space a (purple bar a). A noise vector is then integrated to enhance continuity within the latent space, producing latent space b (purple bar b). Subsequently, deconvolution layers upsample to restore the spatial dimensions. The output of the upper pathway is trained to match the original input BP-HR probability density map (labels = input BP-HR probability density map), ensuring a compressed latent representation is achieved through reconstruction. All parameters in this pathway are optimized via backpropagation. Lower pathway (prediction branch): The daytime BP-HR probability density map is processed using an identical convolutional structure as the upper pathway to extract features and generate latent space c (purple bar c). This latent space c is then concatenated with the full-day Activity and Posture Transition Probabilities…

• 

  Fig. 4. Entropy of overall sample data point bin assignments before and after entropy reduction preprocessing. a Box plots showing the changes in data point entropy for SBP-HR and DBP-HR before and after preprocessing. Blue boxes represent SBP-HR and orange boxes represent DBP-HR. Within each box, the central line indicates the median, and the red plus sign marks the mean. The box edges correspond to the first and third quartiles, while the whiskers extend to ±1.5 times the interquartile range, n = 2874 biologically independent patient samples. The original average entropy values are 2.44 bits (SD 0.20 bits) for HR-SBP and 2.48 bits (SD 0.21 bits) for HR-DBP; the transition average entropy values signiffcantly decrease to 1.55 bits (SD 0.15 bits) and 1.57 bits (SD 0.14 bits), respectively (p < 0.001, one-tailed t-test); the Cropped entropy average entropy values signiffcantly decrease to 0.50 bits (SD 0.25 bits) for HR-SBP and 0.34 bits (SD 0.23 bits) for HR-DBP (p < 0.001, one-tailed t-test). b Scatter plots illustrating the two-dimensional distribution of data point entropy for SBP-HR and DBP-HR across three stages: “original” denotes the raw time series; “transition” represents the…

• 

  Fig. 5. Prediction accuracy of nocturnal hypertension for ABPM-VAE and FCNN models across training, validation, and test sets. The figure displays the performance of the ABPM-VAE and FCNN models on ABPM subsets for predicting nocturnal hypertension. The blue, orange, and green circle lines represent the training, validation, and testing accuracies of the ABPM-VAE model, respectively. Similarly, the blue, orange, and green cross lines represent the corresponding accuracies for the FCNN model. The line plot shows how accuracy changes over the epochs. Since the training and validation subsets originate from the same ABPM recordings, strong performance on these sets alongside poor performance on the test set indicates that the model may have overfitted to individual noise, thereby impairing generalization.

• 

  Fig. 6. ROC curves of ABPM-VAE and ablation models in test data set. a ROC curve for predicting nocturnal systolic hypertension using the ABPM-VAE and ablation screening strategies. Yellow indicates ABPM-VAE; green indicates the FCNN ablation model. b ROC curve for predicting nocturnal diastolic hypertension using the ABPM-VAE and ablation screening strategies. Yellow indicates ABPM-VAE; green indicates the FCNN ablation model. c ROC curve for predicting overall nocturnal hypertension using the ABPM-VAE and ablation screening strategies. Yellow indicates ABPM-VAE; green indicates the FCNN ablation model. d ROC curve for the ABPM-VAE screening strategy showing the impact of thresholds on sensitivity and specificity. The color gradient represents thresholds, with red for lower values and blue for higher values. When the ROC threshold was set to 0.6, the ABPM-VAE model achieved a PPV of 92.12%, a NPV of 55.20%, a sensitivity of 0.73, and a specificity of 0.84. n = 2874 biologically independent patient samples.

• 

  Fig. 7. MSE performance of ABPM-VAE, ablation, and baseline models across training, validation, and test data sets. All boxes are rendered in a uniform light blue hue. The centre line of each box denotes the median, while a red plus sign marks the mean; the mean value is annotated immediately above each box. The upper and lower edges of each box correspond to the first and third quartiles, respectively, and the whiskers extend to ±1.5 times the interquartile range. Light purple dots represent the performance of each method under different test epochs or parameter settings. The ABPM-VAE model achieved the lowest MSE of 77.41 (SD 1.97), (n = 20 epoches), significantly outperforming all other models (p < 0.001, two-tailed t-test). In comparison, the LightGBM model yielded an MSE of 138.49 (SD 10.86), (n = 48 parameter settings), the FCNN ablation model 196.45 (SD 22.15), (n = 20 epoches), and the LSTM model 228.81 (SD 22.92), (n = 20 epoches).

• 

  Fig. 8. Vanilla gradient saliency maps of ABPM-VAE for nighttime BP-HR density maps prediction. Vanilla saliency maps are computed as the absolute input gradients with respect to predicted nighttime BP-HR distributions, normalized for visualization. High-saliency regions (red) form a consistent band aligning closely with the contours and outer boundaries of daytime SBP-HR (left) and DBP-HR (right) joint probability density maps.