Remote patient monitoring
Our client is a Scientific start-up that develops Remote Patient Monitoring (RPM) product for health providers. The company needed to launch MVP in four months and collect user feedback to raise additional venture capital for the commercialization phase. The team decided to test data-driven approaches for health anomaly detecting and health condition trend forecasting. AI-based monitoring would achieve a claimed reduction in nursing observation costs – as one of the critical for many healthcare providers.
The team should run the whole AI development journey from Ideation, Market Intelligence, Data feasibility testing to ML/DL Models development and integration into the main Product in a very tight schedule. Therefore, the company needed a reliable partner with significant expertise in Biosignal processing, DS algorithms, Machine Learning Operations (MLOps), and Product Development lifecycle understanding.
The big problem was on the data layer feasibility. RPM like telehealth, telemedicine, virtual care, and digital health is designed to free healthcare workers from routine tasks, reduce the cost of the care, and unlock access to quality care for many more patients. RPM has to mimic many nurse/doctor functions based on uncertain and subjective data from wearables and smartphones to achieve this.
We had to deal with many RPM-generated data, which were highly sparse, uncertain, and subjective (such as video/text-based patient state detection). Besides that, Remote monitoring deals with the continuous real-time data stream and requires real-time processing and events detection. Physiological data is highly volatile, and it is hard to find out patterns and trends. The data collected with RPM have no direct diagnostic value and require the search and engineering of significant features and models.
We have defined Algorithmic Models requirements for some of the most critical use cases and checked AI feasibility based on data sourcing & available data health check.
During the Research & Development phase, we performed knowledge extraction and data mining to find functional relations of patient medical/health monitoring data and evaluate significant predictive features. The predictive power of features connected with a patient’s medical condition was substantial for effective diagnosing.
Next, we used Deep Learning models of face/voice/speech/text recognition for patient medical/health/well-being conditions self-assessment and fraud detection. Self-assessment enriches the dataset with symptom labels.
Then we used time series analysis for patterns and trends recognition for early detection and forecasting for a medical condition. It allowed effective treatment prescription and operative correction.
Further, we used statistical models for anomaly detection in the data, patterns, and trends. Effective anomaly detection reduces emergency department attendance and readmission rates.
Finally, we integrated Minimum Viable Models into the Product with a robust MLOps architecture supporting the whole Model Lifecycle.
- A list of potential ideas for improving Product Data readiness
- A lot of not exploited data collected from wearables and smartphones
- The main application of the RPM app is online nurse/doctor e-visiting
- AI MVP readiness in four months and prepared to commercial scale
- MVP-proven product hypothesis:
- Diseases knowledge extraction, patterns, and trends recognitions increase the efficiency of diagnosing and treatment programs by 30-40%
- Effective anomaly detection reduces emergency department attendance and readmission rates by up to 50%
- Effective medical condition early detection and forecasting reduce nursing costs up to 40%