Wirtualne narzędzia dedykowane medycynie – wyniki prac B+R (RPMA.01.02.00-14-d767/20) Mazowiecka Jednostka Wdrażania Projektów Unijnych

1. Research objectives and scope The objective of the project was to develop a technology that reduces operational costs and medical staff workload in remote patient monitoring (RPM) through the use of an intelligent virtual assistant — Smart Signal Bot (SSB) — which automatically handles a significant share of communication and incident management generated by telemedical devices. The system was primarily designed for: patients with chronic diseases (ICD-10: I10, E11, J44, I50), elderly patients, patients enrolled in telemonitoring procedures (IoMT – Internet of Medical Things). 2. Clinical problem scale Analysis of telemonitoring system data showed that patients generate four main categories of events: Event type Share of total alerts Missing measurement Very high Threshold exceedance High Symptom reports Medium Technical problems Very high Previously, each of these events required intervention by Contact Center medical staff, resulting in continuous 24/7 operational costs. 3. Technological solution The project developed Smart Signal Bot (SSB), a system based on: Natural Language Processing (NLP) algorithms, deep neural networks, integration with IoMT devices, HL7 FHIR interoperability standard. SSB automatically handles four core categories of events: Category Automated function Missing measurement reminders, dialogue, escalation Technical issues self-repair, guidance, escalation Health symptoms medical interview, AI report, transfer to Contact Center Parameter alarms trend analysis, alerts, dashboards 4. Algorithmic research results 4.1 Tested architectures Three classes of neural networks were evaluated: RNN (LSTM, GRU), CNN (convolutional networks), Transformer (fine-tuned GPT-2). Model Response accuracy Stability LSTM Seq2Seq Low Low CNN Medium High GPT-2 Transformer Highest Highest The fine-tuned GPT-2 model (1.5 billion parameters) proved to be the best conversational model for medical dialogue. 5. Training data The model was trained on: more than 10 MB of medical dialogue data, patient–system Q&A datasets, symptom datasets, technical IoMT data. Training setup: 20 epochs, Transformer architecture with attention mechanism, transfer learning fine-tuning. 6. System performance Metric Result Prediction accuracy for patient condition escalation 89% Reduction of medical staff interventions 35% Alarm response time Reduced by more than 50% Fully automated event handling Dominant 7. Data visualization The system generates: therapy trend charts, vital parameter dashboards, warning and alarm signals, AI-generated reports for medical staff. Each alert contains:[parameter type] – [time] – [value] – [threshold] 8. Clinical relevance Thanks to SSB: patients receive 24/7 contact, medical staff focus only on critical events, telecare costs are significantly reduced, scalable senior care becomes possible. The system is applicable in: cardiology, diabetology, pulmonology, senior care, community psychiatry. 9. Legal status of the technology The developed algorithms are protected by a patent application:P.442425 (WIPO ST.10/C PL442425)Title: Method for optimizing the cost of patient alert handling…Filing date: 30 September 2022