The collaboration between Jiva.ai and Aevice was not just about training a model, it was a project in what it takes to turn real-world healthcare data into something reliable, repeatable, and usable in a clinical context.

Aevice Health’s remote monitoring approach creates rich longitudinal data from patients outside traditional care settings. That data has enormous potential, but it came with the all too often seen realities of healthcare AI; inconsistent collection patterns, missing values, site-to-site variation. The end goal was to make results interpretable enough for teams to trust and act on.

The significance of the project came from bringing several difficult pieces together: compliant international data sharing, real-world respiratory data preparation, rapid AI model development, and practical integration into Aevice’s operating environment. For Jiva.ai, the project became a valuable example of why successful AI depends as much on data readiness, workflow design, and deployment practicality as it does on model training.

The Challenges

One of the first challenges was having enough usable data across the full model development process that reflected the real-world problem. That meant looking closely at whether the right signals were present, whether they were collected consistently, and whether they could be aligned across different time windows. The serving workflow eventually supported separate prediction horizons, which meant the data structure had to remain consistent across multiple model versions.

Another challenge was the multi-site nature of healthcare data. Data collected across different sites can vary in format, timing, completeness, and operational context. Even when the same kind of information is being collected, the way it is captured can differ enough to affect how it should be processed. This made data engineering a central part of the project.

Healthcare compliance also shaped the project. The ability to download and serve trained models locally was important given healthcare environments often require strict control over where data and models live.

Jiva.ai Platform Capabilities

To support the project, the team built tooling and workflows around the full model lifecycle. This included support for K-Fold validation and model selection, helping compare candidate models in a structured way without relying on a single training run.

A lag-data node was also developed so that time-based signals could be transformed into usable model inputs. For Aevice, this mattered because respiratory and patient-monitoring data is not just a snapshot, trends over time can be an important context. Support for model(s) outputs, with local prediction generation and downloadable results was important for the compliance-conscious environments where trained models and patient-related data needed to remain controlled.

On the platform side, Jiva.ai supported rapid AI model creation through pipelines and model selection workflows. Beyond the modelling workflow, the project also required UI, infrastructure, and SIA-assisted development so that complex pipelines could be built and adjusted more easily when needed.

Data Matters

One of the biggest lessons from the Aevice project was that AI quality starts long before model training.Having the right data means more than having a large dataset. It means collecting the correct data points, collecting them consistently, and preserving enough context to understand what the data represents. For respiratory health, that can include physiological signals, environmental context, time-based patterns, and clinically meaningful labels.

Training and validation also depend on this consistency. If data from different sites follows different rules, or if key fields are missing or defined differently, the model development process becomes harder to trust. A major part of the work was therefore about building a full picture from imperfect data: understanding what was available, what was missing, and what processing was needed before the data could be used.

This is one of the areas where visual AI workflows are especially useful. They make data preparation, model training, and model selection easier to inspect and repeat, rather than leaving the process buried in isolated notebooks or one-off scripts.For healthcare use cases, “more data” is not automatically better. The value comes from having relevant, well-understood, properly prepared data that reflects the environment where the AI will actually be used.

Healthcare Data out in the Real-World

In an ideal world, every dataset arrives clean, complete, labelled, and ready for modelling. Real healthcare data rarely works that way.

The project involved non-structured and semi-structured realities: incomplete rows, shifting formats, different collection patterns, and data that needed careful preparation before it could support model development. A large amount of effort went into data engineering and processing, because that is what makes the later AI work possible. This is an important point for anyone building AI in healthcare. The model is only one part of the system. The surrounding workflow, the quality of the data pipeline, and the ability to reproduce decisions matter just as much.

Future Direction

The project showed that the real breakthrough in healthcare AI is not only the model itself. It is the ability to collect the right data, prepare it responsibly, build and refine models quickly, and integrate them into the systems where clinical decisions are actually made.

The main lesson is simple: healthcare AI is not just about building a model. It is about building the process around the model so that data, training, validation, serving, and compliance all work together. The project demonstrates the benefits of working with Jiva.ai in making that process more practical for teams working with real-world clinical data.