An interpretable 1D convolutional neural network for detecting patient-ventilator asynchrony in mechanical ventilation

Quantifiable identification of flow-limited ventilator dyssynchrony with the deformed lung ventilator model

BACKGROUND

Ventilator dyssynchrony (VD) can worsen lung injury and is challenging to detect and quantify due to the complex variability in the dyssynchronous breaths. While machine learning (ML) approaches are useful for automating VD detection from the ventilator waveform data, scalable severity quantification and its association with pathogenesis and ventilator mechanics remain challenging.

OBJECTIVE

We develop a systematic framework to quantify pathophysiological features observed in ventilator waveform signals such that they can be used to create feature-based severity stratification of VD breaths.

METHODS

A mathematical model was developed to represent the pressure and volume waveforms of individual breaths in a feature-based parametric form. Model estimates of respiratory effort strength were used to assess the severity of flow-limited (FL)-VD breaths compared to normal breaths. A total of 93,007 breath waveforms from 13 patients were analyzed.

RESULTS

A novel model-defined continuous severity marker was developed and used to estimate breath phenotypes of FL-VD breaths. The phenotypes had a predictive accuracy of over 97% with respect to the previously developed ML-VD identification algorithm. To understand the incidence of FL-VD breaths and their association with the patient state, these phenotypes were further successfully correlated with ventilator-measured parameters and electronic health records.

CONCLUSION

This work provides a computational pipeline to identify and quantify the severity of FL-VD breaths and paves the way for a large-scale study of VD causes and effects. This approach has direct application to clinical practice and in meaningful knowledge extraction from the ventilator waveform data.

Flow starvation during square-flow assisted ventilation detected by supervised deep learning techniques

BACKGROUND

Flow starvation is a type of patient-ventilator asynchrony that occurs when gas delivery does not fully meet the patients' ventilatory demand due to an insufficient airflow and/or a high inspiratory effort, and it is usually identified by visual inspection of airway pressure waveform. Clinical diagnosis is cumbersome and prone to underdiagnosis, being an opportunity for artificial intelligence. Our objective is to develop a supervised artificial intelligence algorithm for identifying airway pressure deformation during square-flow assisted ventilation and patient-triggered breaths.

METHODS

Multicenter, observational study. Adult critically ill patients under mechanical ventilation > 24 h on square-flow assisted ventilation were included. As the reference, 5 intensive care experts classified airway pressure deformation severity. Convolutional neural network and recurrent neural network models were trained and evaluated using accuracy, precision, recall and F1 score. In a subgroup of patients with esophageal pressure measurement (ΔPes), we analyzed the association between the intensity of the inspiratory effort and the airway pressure deformation.

RESULTS

6428 breaths from 28 patients were analyzed, 42% were classified as having normal-mild, 23% moderate, and 34% severe airway pressure deformation. The accuracy of recurrent neural network algorithm and convolutional neural network were 87.9% [87.6-88.3], and 86.8% [86.6-87.4], respectively. Double triggering appeared in 8.8% of breaths, always in the presence of severe airway pressure deformation. The subgroup analysis demonstrated that 74.4% of breaths classified as severe airway pressure deformation had a ΔPes > 10 cmH2O and 37.2% a ΔPes > 15 cmH2O.

CONCLUSIONS

Recurrent neural network model appears excellent to identify airway pressure deformation due to flow starvation. It could be used as a real-time, 24-h bedside monitoring tool to minimize unrecognized periods of inappropriate patient-ventilator interaction.

The Development, Optimization, and Validation of Four Different Machine Learning Algorithms to Identify Ventilator Dyssynchrony

Invasive mechanical ventilation can worsen lung injury. Ventilator dyssynchrony (VD) may propagate ventilator-induced lung injury (VILI) and is challenging to detect and systematically monitor because each patient takes approximately 25,000 breaths a day yet some types of VD are rare, accounting for less than 1% of all breaths. Therefore, we sought to develop and validate accurate machine learning (ML) algorithms to detect multiple types of VD by leveraging esophageal pressure waveform data to quantify patient effort with airway pressure, flow, and volume data generated during mechanical ventilation, building a computational pipeline to facilitate the study of VD.

MATERIALS AND METHODS

We collected ventilator waveform and esophageal pressure data from 30 patients admitted to the ICU. Esophageal pressure allows the measurement of transpulmonary pressure and patient effort. Waveform data were cleaned, features considered essential to VD detection were calculated, and a set of 10,000 breaths were manually labeled. Four ML algorithms were trained to classify each type of VD: logistic regression, support vector classification, random forest, and XGBoost.

RESULTS

We trained ML models to detect different families and seven types of VD with high sensitivity (>90% and >80%, respectively). Three types of VD remained difficult for ML to classify because of their rarity and lack of sample size. XGBoost classified breaths with increased specificity compared to other ML algorithms.

DISCUSSION

We developed ML models to detect multiple types of VD accurately. The ability to accurately detect multiple VD types addresses one of the significant limitations in understanding the role of VD in affecting patient outcomes.

CONCLUSION

ML models identify multiple types of VD by utilizing esophageal pressure data and airway pressure, flow, and volume waveforms. The development of such computational pipelines will facilitate the identification of VD in a scalable fashion, allowing for the systematic study of VD and its impact on patient outcomes.

Automated evaluation of typical patient-ventilator asynchronies based on lung hysteretic responses

BACKGROUND

Patient-ventilator asynchrony is common during mechanical ventilation (MV) in intensive care unit (ICU), leading to worse MV care outcome. Identification of asynchrony is critical for optimizing MV settings to reduce or eliminate asynchrony, whilst current clinical visual inspection of all typical types of asynchronous breaths is difficult and inefficient. Patient asynchronies create a unique pattern of distortions in hysteresis respiratory behaviours presented in pressure-volume (PV) loop.

METHODS

Identification method based on hysteretic lung mechanics and hysteresis loop analysis is proposed to delineate the resulted changes of lung mechanics in PV loop during asynchronous breathing, offering detection of both its incidence and 7 major types. Performance is tested against clinical patient data with comparison to visual inspection conducted by clinical doctors.

RESULTS

The identification sensitivity and specificity of 11 patients with 500 breaths for each patient are above 89.5% and 96.8% for all 7 types, respectively. The average sensitivity and specificity across all cases are 94.6% and 99.3%, indicating a very good accuracy. The comparison of statistical analysis between identification and human inspection yields the essential same clinical judgement on patient asynchrony status for each patient, potentially leading to the same clinical decision for setting adjustment.

CONCLUSIONS

The overall results validate the accuracy and robustness of the identification method for a bedside monitoring, as well as its ability to provide a quantified metric for clinical decision of ventilator setting. Hence, the method shows its potential to assist a more consistent and objective assessment of asynchrony without undermining the efficacy of the current clinical practice.

A Novel Application of Spectrograms with Machine Learning Can Detect Patient Ventilator Dyssynchrony

Patients in intensive care units are frequently supported by mechanical ventilation. There is increasing awareness of patient-ventilator dyssynchrony (PVD), a mismatch between patient respiratory effort and assistance provided by the ventilator, as a risk factor for infection, narcotic exposure, lung injury, and adverse neurocognitive effects. One of the most injurious consequences of PVD are double cycled (DC) breaths when two breaths are delivered by the ventilator instead of one. Prior efforts to identify PVD have limited efficacy. An automated method to identify PVD, independent of clinician expertise, acumen, or time, would potentially permit early, targeted treatment to avoid further harm. We performed secondary analyses of data from a clinical trial of children with acute respiratory distress syndrome. Waveforms of ventilator flow, airway pressure and esophageal manometry were annotated to identify DC breaths and underlying PVD subtypes. Spectrograms were generated from those waveforms to train Convolutional Neural Network (CNN) models in detecting DC and underlying PVD subtypes: Reverse Trigger (RT) and Inadequate Support (IS). The DC breath detection model yielded AUROC of 0.980, while the multi-target detection model for underlying dyssynchrony yielded AUROC of 0.980 (RT) and 0.976 (IS). When operating at 75% sensitivity, DC breath detection had a number needed to alert (NNA) 1.3 (99% specificity), while underlying PVD had a NNA 1.6 (98.5% specificity) for RT and NNA 4.0 (98.2% specificity) for IS. CNNs using spectrograms of ventilator waveforms can identify DC breaths and detect the underlying PVD for targeted clinical interventions.

Neural Network-Enabled Identification of Weak Inspiratory Efforts during Pressure Support Ventilation Using Ventilator Waveforms

During pressure support ventilation (PSV), excessive assist results in weak inspiratory efforts and promotes diaphragm atrophy and delayed weaning. The aim of this study was to develop a classifier using a neural network to identify weak inspiratory efforts during PSV, based on the ventilator waveforms. Recordings of flow, airway, esophageal and gastric pressures from critically ill patients were used to create an annotated dataset, using data from 37 patients at 2-5 different levels of support, computing the inspiratory time and effort for every breath. The complete dataset was randomly split, and data from 22 patients (45,650 breaths) were used to develop the model. Using a One-Dimensional Convolutional Neural Network, a predictive model was developed to characterize the inspiratory effort of each breath as weak or not, using a threshold of 50 cmH₂O*s/min. The following results were produced by implementing the model on data from 15 different patients (31,343 breaths). The model predicted weak inspiratory efforts with a sensitivity of 88%, specificity of 72%, positive predictive value of 40%, and negative predictive value of 96%. These results provide a 'proof-of-concept' for the ability of such a neural-network based predictive model to facilitate the implementation of personalized assisted ventilation.

Matched Filter Interpretation of CNN Classifiers with Application to HAR

Time series classification is an active research topic due to its wide range of applications and the proliferation of sensory data. Convolutional neural networks (CNNs) are ubiquitous in modern machine learning (ML) models. In this work, we present a matched filter (MF) interpretation of CNN classifiers accompanied by an experimental proof of concept using a carefully developed synthetic dataset. We exploit this interpretation to develop an MF CNN model for time series classification comprising a stack of a Conv1D layer followed by a GlobalMaxPooling layer acting as a typical MF for automated feature extraction and a fully connected layer with softmax activation for computing class probabilities. The presented interpretation enables developing superlight highly accurate classifier models that meet the tight requirements of edge inference. Edge inference is emerging research that addresses the latency, availability, privacy, and connectivity concerns of the commonly deployed cloud inference. The MF-based CNN model has been applied to the sensor-based human activity recognition (HAR) problem due to its significant importance in a broad range of applications. The UCI-HAR, WISDM-AR, and MotionSense datasets are used for model training and testing. The proposed classifier is tested and benchmarked on an android smartphone with average accuracy and F1 scores of 98% and 97%, respectively, which outperforms state-of-the-art HAR methods in terms of classification accuracy and run-time performance. The proposed model size is less than 150 KB, and the average inference time is less than 1 ms. The presented interpretation helps develop a better understanding of CNN operation and decision mechanisms. The proposed model is distinguished from related work by jointly featuring interpretability, high accuracy, and low computational cost, enabling its ready deployment on a wide set of mobile devices for a broad range of applications.

Reconstructing asynchrony for mechanical ventilation using a hysteresis loop virtual patient model

BACKGROUND

Patient-specific lung mechanics during mechanical ventilation (MV) can be identified from measured waveforms of fully ventilated, sedated patients. However, asynchrony due to spontaneous breathing (SB) effort can be common, altering these waveforms and reducing the accuracy of identified, model-based, and patient-specific lung mechanics.

METHODS

Changes in patient-specific lung elastance over a pressure-volume (PV) loop, identified using hysteresis loop analysis (HLA), are used to detect the occurrence of asynchrony and identify its type and pattern. The identified HLA parameters are then combined with a nonlinear mechanics hysteresis loop model (HLM) to extract and reconstruct ventilated waveforms unaffected by asynchronous breaths. Asynchrony magnitude can then be quantified using an energy-dissipation metric, Easyn, comparing PV loop area between model-reconstructed and original, altered asynchronous breathing cycles. Performance is evaluated using both test-lung experimental data with a known ground truth and clinical data from four patients with varying levels of asynchrony.

RESULTS

Root mean square errors for reconstructed PV loops are within 5% for test-lung experimental data, and 10% for over 90% of clinical data. Easyn clearly matches known asynchrony magnitude for experimental data with RMS errors < 4.1%. Clinical data performance shows 57% breaths having Easyn > 50% for Patient 1 and 13% for Patient 2. Patient 3 only presents 20% breaths with Easyn > 10%. Patient 4 has Easyn = 0 for 96% breaths showing accuracy in a case without asynchrony.

CONCLUSIONS

Experimental test-lung validation demonstrates the method's reconstruction accuracy and generality in controlled scenarios. Clinical validation matches direct observations of asynchrony in incidence and quantifies magnitude, including cases without asynchrony, validating its robustness and potential efficacy as a clinical real-time asynchrony monitoring tool.

Interpretable Model Based on Pyramid Scene Parsing Features for Brain Tumor MRI Image Segmentation

Due to the black box model nature of convolutional neural networks, computer-aided diagnosis methods based on depth learning are usually poorly interpretable. Therefore, the diagnosis results obtained by these unexplained methods are difficult to gain the trust of patients and doctors, which limits their application in the medical field. To solve this problem, an interpretable depth learning image segmentation framework is proposed in this paper for processing brain tumor magnetic resonance images. A gradient-based class activation mapping method is introduced into the segmentation model based on pyramid structure to visually explain it. The pyramid structure constructs global context information with features after multiple pooling layers to improve image segmentation performance. Therefore, class activation mapping is used to visualize the features concerned by each layer of pyramid structure and realize the interpretation of PSPNet. After training and testing the model on the public dataset BraTS2018, several sets of visualization results were obtained. By analyzing these visualization results, the effectiveness of pyramid structure in brain tumor segmentation task is proved, and some improvements are made to the structure of pyramid model based on the shortcomings of the model shown in the visualization results. In summary, the interpretable brain tumor image segmentation method proposed in this paper can well explain the role of pyramid structure in brain tumor image segmentation, which provides a certain idea for the application of interpretable method in brain tumor segmentation and has certain practical value for the evaluation and optimization of brain tumor segmentation model.

Multilayer perceptron neural network model development for mechanical ventilator parameters prediction by real time system learning

Background and objective

In pandemic situations like COVID 19, real time monitoring of patient condition and continuous delivery of inspired oxygen can be made possible only through artificial intelligence-based system modeling. Even now manual control of mechanical ventilator parameters is continuing despite the ever-increasing number of patients in critical epidemic conditions. Here a suggestive multi-layer perceptron neural network model is developed to predict the level of inspired oxygen delivered by the mechanical ventilator along with mode and positive end expiratory pressure (PEEP) changes for reducing the effort of health care professionals.

Methods

The artificial neural network model is developed by Python programming using real time data. Parameter identification for model inputs and outputs is done by in corporating consistent real time patient data including periodical arterial blood gas analysis, continuous pulse oximetry readings and mechanical ventilator settings using statistical pairwise analysis using R programming.

Results

Mean square error values and R values of the model are calculated and found to be an average of 0.093 and 0.81 respectively for various data sets. Accuracy loss will be in good fit with validation loss for a comparable number of epochs.

Conclusions

Comparison of the model output is undertaken with physician’s prediction using statistical analysis and shows an accuracy error of 4.11 percentages which is permissible for a good predictive system.

Identifying Patient-Ventilator Asynchrony on a Small Dataset Using Image-Based Transfer Learning

Mechanical ventilation is an essential life-support treatment for patients who cannot breathe independently. Patient-ventilator asynchrony (PVA) occurs when ventilatory support does not match the needs of the patient and is associated with a series of adverse clinical outcomes. Deep learning methods have shown a strong discriminative ability for PVA detection, but they require a large number of annotated data for model training, which hampers their application to this task. We developed a transfer learning architecture based on pretrained convolutional neural networks (CNN) and used it for PVA recognition based on small datasets. The one-dimensional signal was converted to a two-dimensional image, and features were extracted by the CNN using pretrained weights for classification. A partial dropping cross-validation technique was developed to evaluate model performance on small datasets. When using large datasets, the performance of the proposed method was similar to that of non-transfer learning methods. However, when the amount of data was reduced to 1%, the accuracy of transfer learning was approximately 90%, whereas the accuracy of the non-transfer learning was less than 80%. The findings suggest that the proposed transfer learning method can obtain satisfactory accuracies for PVA detection when using small datasets. Such a method can promote the application of deep learning to detect more types of PVA under various ventilation modes.