PaperlayerView exact revision on arXiv

Source-linked AI summary

Detecting and interpreting myocardial infarction using fully convolutional neural networks

Nils Strodthoff, Claas Strodthoff

arXiv:1806.07385v2cs.CYcs.LGstat.ML

TL;DR

The paper addresses myocardial infarction detection from ECG data and the need for interpretable decision criteria. It proposes fully convolutional neural networks evaluated on PTB data, achieving state-of-the-art and human-level performance while revealing clinically recognizable attribution patterns.

  • Problem

    Myocardial infarction detection can benefit from more detailed ECG analysis because laboratory confirmation for NSTEMI may delay treatment.

  • Method

    The paper trains fully convolutional neural networks on 12-lead PTB ECG data and applies attribution methods to examine the network’s decision patterns.

  • Results

    The proposed architecture outperforms state-of-the-art approaches on the PTB dataset and reaches a similar performance level to human cardiologists.

  • Takeaways & Limitations

    Interpretability methods can identify clinically motivated decision rules in ECG attribution maps, supporting assessment of network decisions in medical applications.

  • Takeaways & Limitations

    Attributions are model-dependent and can differ quantitatively or qualitatively across models, while quantitative validation would require annotated ECG data unavailable in PTB.

Abstract

from arXiv · show

Objective: We aim to provide an algorithm for the detection of myocardial infarction that operates directly on ECG data without any preprocessing and to investigate its decision criteria. Approach: We train an ensemble of fully convolutional neural networks on the PTB ECG dataset and apply state-of-the-art attribution methods. Main results: Our classifier reaches 93.3% sensitivity and 89.7% specificity evaluated using 10-fold cross-validation with sampling based on patients. The presented method outperforms state-of-the-art approaches and reaches the performance level of human cardiologists for detection of myocardial infarction. We are able to discriminate channel-specific regions that contribute most significantly to the neural network's decision. Interestingly, the network's decision is influenced by signs also recognized by human cardiologists as indicative of myocardial infarction. Significance: Our results demonstrate the high prospects of algorithmic ECG analysis for future clinical applications considering both its quantitative performance as well as the possibility of assessing decision criteria on a per-example basis, which enhances the comprehensibility of the approach.

I. INTRODUCTION

Myocardial infarction requires timely detection, yet conventional ECG classification often depends on non-trivial preprocessing and hand-engineered features. The paper therefore proposes a fully convolutional approach for direct ECG analysis and investigates whether its decisions correspond to clinically recognized signs.

  • Clinical motivation: Early myocardial infarction detection matters because delayed diagnosis can postpone effective treatment, especially for NSTEMI cases requiring laboratory confirmation.ECG analysis could help accelerate this process.
  • Clinical motivation: Automated ECG algorithms could improve diagnostic accuracy, speed, and economic efficiency where cardiologists are unavailable.The introduction also notes that such systems could be tuned to specific decision boundaries.
  • Existing approaches: Conventional ECG classifiers typically preprocess signals, segment heartbeats, extract engineered features, and then apply a classifier.These preprocessing and feature-extraction steps are described as non-trivial.
  • Proposed direction: The paper applies deep neural-network methods to ECG time-series classification, focusing on raw medical-monitoring data such as ECG or EEG.The methodology is framed within broader time-series classification.
  • Contributions: Its contributions include a fully convolutional network for 12-lead PTB ECGs, subdiagnosis and channel analyses, and attribution-based investigation of decision patterns.The paper compares performance with prior approaches and reported cardiologist performance.

II. RELATED WORKS

Prior ECG infarction classifiers commonly rely on preprocessing and extracted features, while this study emphasizes raw multi-channel PTB ECG data as a more realistic benchmark. The dataset is curated to support healthy-versus-infarction classification and selected localization analyses.

  • Research landscape: Time-series classification research includes recurrent and convolutional neural networks, while ECG work commonly targets arrhythmia or myocardial infarction.The paper narrows its related-work discussion to methods relevant to ECG classification.
  • Existing ECG methods: Many myocardial infarction algorithms use classical classifiers after preprocessing and feature extraction, whereas some neural methods operate on beat-level signals or raw ECG data.The paper distinguishes its approach from methods using neural networks only after engineered features or separated beats.
  • Dataset choice: The study chooses the PTB Diagnostic ECG Database because common time-series benchmarks are often short, non-periodic, and single-channel.PTB better matches the paper’s focus on complex, multi-channel medical signals.
  • Dataset construction: The selected PTB data comprise 549 records from 290 subjects, restricted to healthy-control and myocardial-infarction diagnoses.The study excludes records with unknown localization and applies further selection decisions for infarction records.
  • Dataset construction: To limit repeated-patient and post-intervention effects, the study keeps all healthy records but only the first myocardial-infarction ECG from each patient.For records with metadata, infarction age has a median of 2.0 days and 14% were recorded after intervention.
  • Medical labeling: The study groups infarction localization into anterior and inferior classes because the dataset is too small for reliable discrimination across all detailed subclasses.These subdiagnoses are also examined in relation to classification performance.
  • Signal representation: The analysis uses eight clinically relevant leads after discarding redundant limb leads and rarely used Frank leads.The PTB database provides 15 simultaneously measured channels before this reduction.

IV. CLASSIFYING ECG USING DEEP NEURAL NETWORKS

The classification procedure addresses long ECG sequences, class imbalance, metric choice, and uncertainty estimation. It uses fixed windows, minority-class oversampling, Youden’s J-statistic for model selection, and patient-level 10-fold cross-validation.

  • Input handling: Long ECG time series require either attention mechanisms or extraction of random subsequences before neural-network classification.The selected procedure leaves window size and optional downsampling as its key hyperparameters.
  • Input handling: The study uses 4-second windows downsampled to 192 input points, large enough to capture at least three normal-rate heartbeats.These settings were fixed because the dataset was too small for extensive hyperparameter optimization.
  • Class balance: The dataset contains 80 healthy records and 127 myocardial-infarction records, so the minority healthy class is oversampled at a 2:1 ratio.This procedure addresses the binary classification imbalance.
  • Model selection: Youden’s J-statistic is used for model selection because sensitivity and specificity are clinically common metrics and accuracy depends on class proportions.J is defined as sensitivity + specificity − 1.
  • Evaluation metrics: Sensitivity and specificity are distinguished from other observables such as F1 and F2 scores, which combine precision-related quantities with sensitivity.The passage identifies true-positive, false-negative, and false-positive outcomes in the J-statistic formulation.
  • Evaluation procedure: The study performs 10-fold cross-validation to estimate test sensitivity and specificity, while cautioning that small-data fluctuations do not directly provide estimator variance.The authors note that variance cannot be inferred from cross-validation scores by simple means.

B. Investigated architectures

The paper compares convolutional and recurrent architectures for ECG time series, including fully convolutional and residual-style networks. It also tests frequency-domain inputs while retaining a direct-input design without preprocessing.

  • Architecture families: The study investigates convolutional and recurrent neural-network architectures for time-series classification.Recurrent models are included because they are a natural choice for sequential data.
  • Convolutional models: Its convolutional variants are inspired by fully convolutional networks and residual networks used successfully in computer vision.Architectural details are provided separately in the appendix.
  • Input representations: The models can operate directly on downsampled time-series data or incorporate frequency-domain inputs obtained with a Fourier transform.The frequency-domain experiments leave the network architectures otherwise unchanged.
  • Input representations: The approach operates directly on the downsampled ECG input without preprocessing steps.This distinguishes it from methods that require signal preprocessing before neural classification.
  • Recurrent models: The recurrent comparison uses LSTM cells with either a final-output fully connected layer or an additional time-distributed dense layer.The second variant adds a time-series classification objective with adjusted loss functions.

V. RESULTS

On 12-lead data, convolutional architectures performed competitively, with fully convolutional and ResNet models achieving similar time-domain performance and outperforming recurrent architectures.

  • Evaluation: Classification results are reported using cross-validated J-statistics for different network architectures on 12-lead data.The study does not report confidence intervals because test sets contain 20 or fewer patients.
  • Network architecture comparison: Fully convolutional and ResNet architectures achieve similar performance on time-domain data.The comparison does not establish a ranking between the two convolutional architectures on this dataset.
  • Network architecture comparison: Convolutional architectures show a better score than recurrent architectures.The recurrent comparison used standard LSTMs and did not investigate more advanced mechanisms such as attention.
  • Selected architecture: The fully convolutional architecture is used for subsequent analyses because it matches the best-performing ResNet with a simpler architecture.The default setup for later investigations uses 12-lead data.

B. MI localization, benchmarks, and channel selection

The study examines myocardial-infarction subdiagnoses using fully convolutional models, finding different difficulty levels across anterior and inferior infarction and evidence that the model separates medically motivated subclasses.

  • Subdiagnosis performance: Inferior myocardial infarction classification performs worse than anterior myocardial infarction classification across training procedures.The authors attribute this difference to more typical signs in Wilson leads for anterior infarctions.
  • Subdiagnosis performance: Generic myocardial-infarction training improves inferior-infarction classification, whereas specific anterior-infarction training performs better for the simpler anterior task.The comparison concerns models trained with generic or subdiagnosis-specific labels.
  • Training procedure: The model trained without subdiagnoses achieves slightly higher scores for both unspecific infarction and aMI/iMI classification.The authors suggest insufficient training data might explain this result and restrict subsequent investigations to that model.
  • MI localization and subclass discrimination: The subdiagnosis confusion matrix indicates that the model discriminates between aggregated aMI and iMI classes.This supports the medical motivation for assigning the aggregated subdiagnoses.
  • Human comparison: Subdiagnosis evaluation enables comparison with human cardiologist performance, supporting an at-least-human-level performance claim for the task.The comparison is limited by different datasets, performance metrics, and subdiagnosis assignments; the authors reject a superhuman-performance claim.

2) Comparison to literature approaches:

The literature comparison emphasizes patient-level cross-validation and evaluates the fully convolutional approach under a closely replicated setup, where it outperforms the compared literature results while operating on input data without preprocessing filters.

  • Benchmark criteria: Patient-level cross-validation is emphasized because many literature results either omit cross-validated scores or risk data leakage.The authors restrict comparison to recent works reporting patient-level sampling and cross-validated scores.
  • Comparison to literature approaches: Under a closely replicated literature setup, the approach shows superior performance to the compared results.The replicated setup includes limb leads and healthy records alongside genuine inferior myocardial-infarction ECGs.
  • Comparison to literature approaches: Unlike the compared literature algorithms, the approach operates directly on input data without preprocessing with appropriate input filters.This comparison uses the fully convolutional architecture.
  • Benchmark criteria: The authors regard the 12-lead benchmark results as more suitable because clinical ECGs use the full set of eight non-redundant channels.They also identify using the first infarction ECG per patient as better suited to acute-infarction classification.

3) Channel selection:

Channel combinations reveal how much information different ECG leads contribute to myocardial infarction classification. Performance improves when limb leads are complemented with Wilson leads, with the best overall result using all channels.

  • Channel combinations estimate the relative information contributed by each lead to the classification decision.
  • Lead III provides the least information among leads I, II, and III, possibly because its direction poorly matches the heart’s usual electrical axis.
  • Adding Wilson leads to limb leads further improves classification performance toward the standard 12-lead setup.
  • The overall best result uses all channels, although this does not correspond to the clinically relevant setting in which conventionally only 12 leads are available.

C. Interpretability

The study applies attribution methods to examine which ECG regions and channels influence convolutional-network decisions. Attributions often align with clinically recognized infarction signs, but they remain model-dependent and may highlight surrounding regions rather than the exact clinical feature.

  • Attribution methods are used qualitatively to investigate whether the model relies on features similar to those recognized by human cardiologists.
  • Attributions remain consistent across beats within an ECG and show reasonable consistency for similar ECG features across patients.
  • Highlighted regions include established infarction signs such as ST-segment elevation, pathological Q waves, and T-wave inversion.
  • A convolutional network may attribute importance to the QRS complex and surrounding regions when detecting ST elevation because it compares information before and after the QRS complex.
  • The model emphasizes Wilson leads for anterior infarction and limb leads for inferior infarction, matching where clinicians expect infarction signs.
  • Attributions can differ quantitatively and qualitatively across models, including models with different random initializations.

VI. SUMMARY AND CONCLUSIONS

The study presents a fully convolutional neural network for myocardial infarction detection on 12-lead PTB ECG data, with performance comparable to human cardiologists and exploratory interpretability analysis. Larger databases are needed for quantitative precision and broader diagnostic coverage.

  • The proposed network detects myocardial infarction on clinically relevant 12-lead PTB ECG data and reaches a similar performance level to human cardiologists.The study also emphasizes careful data selection and cross-validation.
  • Interpretability methods identified common cardiologists’ decision rules in the network’s attribution maps.The analysis is presented as an exploratory step toward understanding network decisions in medical applications.
  • Larger databases are required for quantitative precision in attribution analysis and further classification improvements.The authors also connect larger datasets with extending the method to additional subdiagnoses and cardiac conditions.

APPENDIX

The appendix describes the model implementations and alternative architectures, including fully convolutional, ResNet-inspired, and recurrent designs. It also specifies input representation, normalization, optimization, frequency-domain inputs, and training details.

  • All models use input-channel batch normalization as the only preprocessing step and minimize crossentropy loss with Adam at learning rate 0.001.The models were implemented in TensorFlow.
  • ECG signals are represented as two-dimensional tensors of sampling points by input channels, enabling one-dimensional convolutional and pooling operations.This contrasts with the three-dimensional tensor representation conventionally used for images.
  • Convolutional and fully connected layers generally use ELU activations, while architectures with fully connected layers apply dropout at rate 0.5.The appendix reports a slight performance increase compared with standard ReLU activation with batch normalization.
  • Convolutional architectures: The investigated convolutional models include a fully convolutional architecture with global average pooling and a ResNet-inspired architecture with skip connections.The study also evaluates frequency information from a Fast Fourier Transformation using NFFT = 2^⌈log2(d)⌉, with d = 192 after rescaling.
  • Recurrent architectures: The recurrent alternatives use LSTM cells, including a 256-hidden-unit single-cell design with either final-output or time-distributed fully connected layers.Stacked LSTM architectures provided no significant performance gain, and gradient clipping was applied during training.
  • Convolutional architectures: Figure 3 contrasts the fully convolutional and ResNet convolutional architectures.
Loading 1806.07385v2…