Modeling Indonesian Heart Sound Signals
Reflections on the workflow behind turning raw sensor recordings into a usable, published research dataset.
Research in applied engineering rarely starts with a clean dataset — it starts with raw recordings, inconsistent labeling, and a question worth answering. Modeling a dataset of Indonesian heart sound signals, published at IC2IE 2024, followed that pattern closely.
Why the dataset mattered
Heart sound (phonocardiogram) research depends heavily on the datasets available to train and validate detection models. Most widely used public datasets weren't collected in Indonesian clinical settings, which limits how well models trained on them generalize locally. Building and modeling a locally sourced dataset addresses that gap directly.
The general workflow
- Signal acquisition — capturing heart sound recordings under consistent conditions to minimize noise variance between samples.
- Preprocessing — filtering, normalization, and segmentation so recordings are comparable across subjects.
- Labeling — structuring recordings against known heart sound categories with domain input.
- Modeling the dataset — documenting statistical properties and structure so the dataset itself is usable by other researchers, not just by one model.
What's easy to underestimate
- Documentation is part of the deliverable. A dataset without clear structure and metadata is far less valuable to the research community than a smaller, well-documented one.
- Cross-disciplinary review matters. Engineering rigor on signal processing has to be checked against clinical/domain expectations for the labeling to hold up.
- Publication is a milestone, not an endpoint. A published dataset is only useful if others can actually get it and use it correctly.
Working across engineering and applied research — rather than treating them as separate tracks — is what made it possible to take this from a sensor-and-signal problem to something publishable and reusable.