Models today are getting bigger and smarter, surpassing new benchmarks every other day. Yet, for all their sophistication, most of the focus is on achieving peak performance on pristine, well-behaved datasets. This often makes us forget one equally important factor: noise. Real-world observations are inherently noisy and don't always adhere to perfect benchmarks, and this is one of the issues we try to tackle through this work. This is a crucial issue, especially in medicine, and it's precisely what we've tackled through our latest work on arrhythmia detection from heart sound recordings.

The Problem: When Real-World Noise Obscures a Life-Saving Diagnosis

Cardiovascular diseases (CVDs) remain the foremost cause of mortality worldwide. The early detection of irregularities like heart arrhythmias (rhythms that are too fast, too slow, or erratic) is crucial to preventing severe complications such as strokes, heart failure, and cardiac arrest.

While deep learning has emerged as a powerful tool to automate the diagnosis of heart conditions from audio recordings—specifically phonocardiograms (PCG)—existing models often struggle to generalize when faced with the messy reality of clinical data.

The challenges, as seen in the PhysioNet CinC Challenge 2016 dataset, are threefold:

1. Noise and Artifacts: Real heart sound recordings are full of noise components and high-frequency artifacts that can drastically impair the performance of arrhythmia detection models.
2. Small and Imbalanced Datasets: Biomedical data is often limited, and the distribution is severely skewed. The dataset we used shows a notable skew, with roughly 87% (5154 samples) classified as normal, and a minority (771 samples) as abnormal heart sounds.
3. Generalization: Current models often struggle to generalize well in real-world scenarios, especially when dealing with small or noisy datasets.

We set out to build an architecture that is not just accurate but also robust to these real-world challenges.

An Adaptive Memory Inspired by Control Theory - The H$_{\infty}$ Filter Enhanced CNN-LSTM

Before any deep model sees the data, we apply a dual-stage denoising pipeline to each heart sound. First, a discrete wavelet transform breaks the audio into multi-scale frequency components, capturing the rhythmic structure across time scales. Next, a low-pass IIR (Infinite Impulse Response) filter smooths the wavelet result, suppressing high-frequency artifacts that usually carry noise, which are later converted into Mel Spectrograms. Our solution is a novel deep learning architecture called the CNN-H$_{\infty}$-LSTM, which processes Mel spectrograms of heart rhythm audio. The CNN layers identify spatial features, while the LSTM layers track temporal dependencies.

Re-engineering the LSTM Cell: H$_{\infty}$ Filter Integration

We substituted the default forget gate of the LSTM with a parameterized mechanism inspired by the H-Infinity ($H_{\infty}$) filter from control theory, creating a new recurrent unit called the $H_{\infty}$ LSTM cell.

The $H_{\infty}$ filter is designed to provide a guaranteed bound on estimation error even under unknown-but-bounded disturbances and model inaccuracies. By passing a trainable parameter $K_{filter}$ through a sigmoid activation to derive a robustness coefficient, $\lambda_{h}$ ], our model can dynamically control the trade-off between retaining past memory and incorporating new information. This allows the model to learn a more robust, data-driven forgetting mechanism compared to the gradient-based updates in standard LSTMs

Our architecture combines the strengths of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs).

• CNN for Spatial Feature Extraction: The spectrogram is first processed by a series of convolutional blocks. Each block contains two 3x3 convolutional layers with "same" padding, followed by Batch Normalization to stabilize training and a 2x2 MaxPooling layer to downsample the feature maps.

• The H$_{\infty}$-LSTM for Temporal Analysis: The core innovation lies in replacing the standard LSTM forget gate with a mechanism inspired by the H-Infinity ($H_{\infty}$) filter. While a normal LSTM computes its cell state as $c_{t}=f_{t}c_{t-1}+i_{t}\tilde{c}_{t}$, our version replaces the volatile forget gate ($f_t$) with a stable, learnable robustness coefficient ($\lambda_{h}$):

  $c_{t}=(1-\lambda_{h})c_{t-1}+\lambda_{h}i_{t}\tilde{c}_{t}$

  This provides a guaranteed bound on estimation error, allowing the model's memory to remain stable even when processing noisy, unpredictable inputs.

Tackling Class Imbalance

To address the extreme skew in the data, we introduced a custom training strategy:

• Penalty-Weighted Loss (PWL): This custom loss function dynamically adjusts the loss based on the model's misclassification behavior, with a particular focus on false negatives and false positives, which are critical in medical diagnosis.
• Stochastic Adaptive Probe Thresholding (SAPT): This adaptive framework dynamically learns the optimal decision threshold $\tau^{*}$ during training to maximize a task-specific metric like the F1-score. This prevents the model from being biased towards the majority class.

Real-World Impact: High-Performance, Accessible Diagnostics

The proposed $\text{CNN-H}_{\infty}\text{-LSTM}$ model achieved state-of-the-art results on the benchmark PhysioNet CinC Challenge 2016 dataset, demonstrating superior performance over pre-trained audio and vision models, as well as existing approaches.

Why the leap? The H∞ filtering makes the temporal model highly robust to real-world distortions, and the custom training loss/threshold ensure sensitivity to the rare positives. Together, the network isn’t just memorizing “clean” samples; it’s learning to ignore spurious noise. This high level of performance means a more reliable system that can significantly change how we approach cardiac screening:

• Robustness in Clinics: The integration of the $H_{\infty}$ filter improves robustness against noise and variability, leading to a higher degree of generalization on small and noisy datasets. This is crucial for clinical settings.
• Scalable and Low-Cost Screening: The end-to-end design enables real-time deployment on mobile or edge devices, supporting scalable and low-cost cardiac screening.

By prioritizing robustness and intelligently handling the imperfections of real-world data, the $\text{CNN-H}_{\infty}\text{-LSTM}$ architecture takes us a critical step closer to delivering scalable, highly reliable, and life-saving cardiac screening to a global population.

Future Work

Modern ML must move beyond chasing leaderboard scores on pristine data. Especially in healthcare and edge computing, we deal with low-power devices and messy signals. The H∞-augmented LSTM is an example of bringing control-theoretic robustness into deep nets, acting like an automatic noise filter built into the model.

Looking ahead, such techniques – robust filtering, adaptive thresholds, loss shaping – should become more mainstream as ML expands into the wild and we move closer to ML systems that work in practice, not just in theory