MODULUE
CASING
CKT
MIC
RECIVER
TRANSMITTER

About the Project, RYTHM.IX

Inspiration

Cardiovascular diseases remain one of the leading causes of death globally, yet regular screening is often skipped due to time, cost, limited access to specialists, and post-pandemic changes in healthcare-seeking behavior. We were motivated to bridge this gap with a solution that:

Empowers anyone to check heart health at home, without medical expertise
Uses AI to detect early signs of cardiac abnormalities
Reduces unnecessary specialist visits while encouraging preventive care

Our vision: make proactive heart health monitoring as simple as recording a short audio clip.

What We Built

RYTHM.IX is an AI-powered heart sound classification system with:

A user-friendly web dashboard for recording/uploading heart sounds
Real-time oscillogram visualization (amplitude vs. time)
Deep learning–based classification (e.g., Normal, Murmur, Artifact, Extrahls, Extrastole)
Actionable summaries with confidence scores
Secure data handling with end-to-end encryption and AES-128 at rest
Cloud-backed processing for scalability

Conceptual pipeline:

Capture heart sound via a digital stethoscope or microphone
Denoise and segment audio; extract relevant features
Run inference with a TensorFlow model
Display predictions, waveform plots, and recommendations

How We Built It

Frontend: Responsive React-based dashboard for recording, uploads, and visualization
Audio: Client-side capture and waveform rendering; server-side preprocessing (filtering, segmentation)
ML: TensorFlow deep learning classifier trained on labeled heart sound patterns
Features: Time–frequency representations (e.g., spectrogram/MFCC variants) and engineered temporal cues
Backend: REST endpoints for upload, inference, and results retrieval
Cloud: Scalable model serving and storage
Security: TLS for data in transit; AES-128 for data at rest; user authentication and controlled access

A typical classification head uses a softmax layer; the confidence for class i is:

Inline: $$ P(\text{class}i \mid x) = \dfrac{e^{f_i(x)}}{\sum{j=1}^{K} e^{f_j(x)}} $$
Display: $$ P(\text{class}i \mid x) = \frac{e^{f_i(x)}}{\sum{j=1}^{K} e^{f_j(x)}} $$ where $$ f_i(x) $$ is the model logit for input features $$ x $$, and $$ K $$ is the number of classes.

What We Learned

Signal processing matters: robust denoising and segmentation dramatically improve model stability
Feature design + DL: combining time–frequency features with CNNs yields better generalization than raw waveforms alone in noisy conditions
UX for healthcare: clear visualizations and simple language are essential for non-expert users
Security by design: encryption, consent, and data retention choices must be first-class features
Model evaluation: stratified validation and environment-like noise testing are critical to avoid optimistic accuracy

Challenges We Faced

Noisy real-world recordings: Motion artifacts, breathing noise, clothing friction required careful preprocessing
Class imbalance: Some conditions (e.g., extrastole) were underrepresented; we used augmentation and weighted losses
Thresholding and triage UX: Balancing sensitivity (early detection) and specificity (avoiding false alarms) for user trust
Latency: Keeping total round-trip inference within seconds on modest networks
Regulatory alignment: Designing with medical compliance and future certification in mind

Technical Highlights

Multi-class classifier for Normal, Murmur, Artifact, Extrahls, Extrastole
Oscillogram and time-series visualization for 2–10 second windows
Confidence-aware recommendations and retest prompts for low-SNR inputs
Modular architecture to plug in improved models without UI changes

Confidence-weighted decision suggestion:

Inline: $$ \text{suggestion} = \arg\max_i \left( \alpha \, P(\text{class}_i \mid x) + \beta \, q_i \right) $$
Display: $$ \text{suggestion} = \arg\max_i \left( \alpha \, P(\text{class}_i \mid x) + \beta \, q_i \right) $$ where $$ q_i $$ encodes class-specific clinical priority heuristics, and $$ \alpha, \beta $$ balance confidence vs. priority.

Impact

Encourages regular, at-home screening for at-risk groups (elderly, cardiac patients, rural/remote users)
Reduces routine clinical load by triaging normal vs. potentially abnormal cases
Promotes preventive care and early intervention

Next Steps

Mobile app with offline capture and later sync
Model upgrades with broader datasets and domain adaptation
Clinician portal for longitudinal monitoring
Multi-language support and accessibility improvements
Compliance pathway (e.g., clinical validation, certifications)

Acknowledgements

Built with a commitment to accessible health technology and inspired by the need to make preventive cardiac care practical, private, and proactive for everyone.