Enhancing BP Prediction Performance with Model Agnostic Meta Learning

The traditional method of blood pressure monitoring using a sphygmomanometer involves a cuff which is obtrusive and can be inconvenient for repeated measurements. Recent advancements in wearable technology and photoplethysmography (PPG) have introduced a cuff-less approach for blood pressure measurement. MFine has developed a mobile app that utilizes PPG to measure blood pressure. The PPG signal extraction, feature engineering and machine learning steps employed in MFine App is reported in our previous article.

The limitation of the feature engineering and machine learning approach reported in the previous article are:

1. The feature engineering approach are less robust as these features are hand-crafted and can be affected by the frames per second (FPS) of the recorded video. The contraction and expansion of heart is responsible for blood pressure and these events occur sequentially. Therefore, sequence model such as Long Short Term Memory (LSTM) can be trained to learn robust features.
2. The feature engineering and machine learning model achieved the best performance for the normal range of BP values, approximately 80/120. (refer here). However this performance difference is attributed to the non-uniform dataset, where the samples representing hypertension and hypotension BP are limited in number compared to the samples within the normal range.

To address the performance limitation due to hand crafted features and the non-uniform data distribution, we pursued two approaches:

1. Data Collection: We conducted an internal data collection drive wherein we collected ground truth BP values and the PPG signal before and after certain physical activity. The activity involved climbing three floors of MFine office using stairs. This data collection drive led to an improved data distribution.
2. Few Shot Learning: The few shot learning approach allows the model weights to be updated using a smaller number of samples of newer class data.

The article discusses the architecture of the LSTM model and the implementation of few shot learning method to enhance the prediction performance of BP.

LSTM Model for BP Prediction

Sequence models in the deep learning framework have the ability to integrate features that remain unaffected by changes in the sampling rate and are closely connected to the underlying physiological process. The proposed LSTM-based blood pressure (BP) prediction method has been accepted at EUSIPCO 2023, and the full paper can be accessed here.

Figure 1. Proposed BP measurement method and the LSTM architecture.

The proposed LSTM model in Figure 1 is trained using 1,300 PPG samples collected using 18 different android smartphones. The dataset was obtained from walk-in patients at a hospital resulting in a large number of normal BP samples and smaller number of hypertension BP samples. The trained model be defined using the parameters θ and the mode function is defined as F_θ.

Model Agnostic Meta Learning for BP Prediction

Recent research has been focused on enhancing machine learning algorithms to make them adaptable to different tasks. Few-shot learning methods have emerged as a solution to learn new tasks with minimal data. One such approach, model-agnostic meta-learning (MAML), has gained significant attention in the context of regression problems.

The implementation of MAML we discuss here is based on the robotic domain. Let’s consider a scenario where a robotic arm is trained to grab objects such as plastic bottles using a large dataset. However, if the dataset contains fewer samples for grabbing tin cans, glass bottles, paper cups, and so on, MAML can be implemented to learn new weights using a smaller number of samples. In MAML, the samples from different tasks are treated as training instances.

In our work, we have implemented MAML as a few-shot learning method. We include a dataset with hypertension blood pressure (BP) values as part of the support and query datasets. The support and query datasets are defined according to Equation 1. During the meta-training phase, gradients are computed based on the support set data, and the pre-trained weights are updated accordingly. Subsequently, during the meta-testing phase, gradients are computed on the query set, and the weights are updated following Equation 1.

Equation 1. MAML

During the internal data collection drive, approximately 1000 PPG samples were collected. These samples were divided into a support set, consisting of 700 samples, and a query set, consisting of 300 samples. The MAML implementation was then applied to the pre-trained LSTM model discussed in the previous section.

Results

We conducted a comparison of the performance of three different models, namely model-1, model-2, and model-3, using the same test dataset. Model-1 was trained using 1300 PPG samples collected from walk-in patients at the hospital, and its block diagram is depicted in Figure 1. For model-2, we re-trained model-1 using the 700 PPG samples obtained from the internal data collection drive, which included physical activity. Model-3 utilized the MAML approach, where the weights of the pre-trained model-1 were updated using the 700 PPG samples collected internally. The performance of all the three model is assessed using 200 PPG samples, referred to as test dataset.

The overall blood pressure (BP) prediction performance of model-1, model-2, and model-3, in terms of Mean Absolute Error (MAE), is presented in Table 1.

Table 1. Comparison of BP prediction performance

It is evident from the results that the MAML approach yields superior performance compared to the other two models. Furthermore, Figure 2 illustrates the Absolute Error (AE) for blood pressure (BP) predictions across the entire range of BP values for model-1, model-2, and model-3.

Figure 2. Absolute error of BP prediction of model-1, 2, and 3.

Ideally, the Absolute Error (AE) should remain flat for all blood pressure (BP) values. However, due to the large number of samples representing normal BP values (i.e., 80/120 mmHg), the AE curve exhibits a V-shaped pattern with the least error occurring at 80/120. Model-1 replicates this V-shaped characteristic in its AE curve.

The internally collected dataset, which includes physical activity, has a slightly higher number of hypertension samples compared to the original dataset. By retraining the model using this new dataset, model-2 achieves the least AE for a wider range of BP values, as indicated in the middle plot of Figure 2. Due to the physical activity, there is an increase in the number of hypertension samples, while the hypotension samples remain the same. Consequently, for model-2, the least AE is observed at 80 for Diastole and 120–130 for Systole. This suggests that having a more uniform dataset with an equal number of samples representing different BP values can improve the V-shaped AE curve and transform it into a U-shaped curve.

Model-3, which utilizes the few-shot learning approach implemented with MAML, achieves better performance compared to model-1 and model-2. The acceptable range of error according to American Association of Measurement Institute (AAMI) 5 ± 8 mmHg. The model-1 meets the AAMI criteria for systole range of 115 to 120 mmHg whereas, model-3 meets the criteria for a wider systole range of 115 to 135 mmHg. The physical activity resulted in higher systole value whereas, the diastole values are less affected. Therefore, all the three models achieve similar performance for diastole BP values.

Concluding Remarks

In conclusion, this research delved into the exciting realm of blood pressure (BP) prediction using cutting-edge deep learning techniques. By harnessing wearable technology and photoplethysmography (PPG), a novel cuff-less approach was introduced, providing a more convenient and hassle-free method for monitoring BP.

The limitations of traditional hand-crafted features were overcome by leveraging the power of sequence models, specifically the Long Short-Term Memory (LSTM) model. Additionally, the introduction of model-agnostic meta-learning (MAML) methods provided a remarkable capability to learn new tasks with minimal data. Model-3, incorporating the MAML approach, emerged as a standout performer, surpassing both model-1 and model-2. The AE analysis uncovered intriguing patterns, revealing that a more balanced dataset representing a wider range of BP values could potentially transform the V-shaped AE curve into a captivating U-shaped curve.

The implications of this research are immense. Imagine a future where individuals can effortlessly monitor their BP using smartphone App, leveraging state-of-the-art deep learning algorithms to accurately predict and manage their health. The potential for advancements in personalized healthcare and preventive medicine is incredibly promising.

As the journey to revolutionize BP monitoring continues, further investigations are warranted. Exploring larger datasets, refining model architectures, and pushing the boundaries of deep learning hold the key to unlocking even greater accuracy and precision in BP prediction. The possibilities are endless, and the path of exploration has only just begun.