The Croatian team Insulogic combined clinical knowledge, health data and an AI voice assistant to identify patients at risk of abandoning diabetes treatment and contact them before they disengage from care.
Medication adherence remains one of the most stubborn problems in chronic disease management. A treatment can be clinically effective, affordable and widely available, yet its benefits depend on patients taking it consistently.
Doctors often discover that treatment has been interrupted only after the consequences become visible. A prescription is collected late, an appointment is missed or a patient returns with poorer disease control. By then, an opportunity for earlier support may already have passed.
Insulogic, the winning team of the 2026 AI4Health.Cro Innovation Challenge, developed a system intended to detect those warning signs sooner and connect prediction with a practical response.
The platform combines a machine-learning model with an AI voice assistant that contacts patients by telephone in Croatian. When the model identifies a patient at risk of interrupting treatment, the assistant calls to ask whether the medication has been taken and whether the patient has encountered difficulties, including side effects or practical barriers.
The conversation is transcribed and transferred to a dashboard where the doctor can review patients considered at risk and see the issues raised during each call. The system is designed to preserve contact between consultations while limiting the routine administrative work placed on clinicians.
Finding signals in patterns of behaviour
The Insulogic team developed its predictive model using 74 variables drawn from patients’ medical histories.
The analysis pointed towards a finding with broad relevance for health data research: patterns of behaviour may carry stronger predictive information than individual clinical measurements.
The regularity of prescription collection, the time between medical visits and other indications of engagement with care appeared particularly useful in identifying patients who might discontinue treatment.
Medical records are rich in diagnoses, prescriptions and laboratory values. They often provide a less direct account of what happens between appointments. Behaviour emerges through sequences: when a patient returns, how consistently prescriptions are collected and whether established routines begin to change. These patterns may reveal disengagement before it appears as an explicit clinical event.
For Insulogic, the competition demonstrated that some of the most useful information in a health dataset lies in the relationship between recorded events. The timing and recurrence of those events can describe a patient’s interaction with the health system more effectively than a single field in the medical record.
The team also focused on the step that follows prediction. A risk score has limited clinical value when it remains one more item for a doctor to interpret within an already crowded workflow.
Insulogic therefore linked the model to an intervention. The platform identifies a potential problem, contacts the patient and returns relevant information to the clinician.
Building across clinical and technical disciplines
Insulogic was developed by a five-member team combining medicine and software development. Dr Adrian Sallabi and Dr Viktor Ivanić, physicians working in emergency care, brought experience of clinical workflows and the consequences of poor adherence. Chiara Krtak, a final-year medical student, helped connect medical knowledge with the interpretation of health data. Josip Alpeza and Hrvoje Hrvoj, co-owners of the development agency Differo Studio, led the technical work.
The team structure shaped the development process. The engineers extracted and processed information from an OMOP-formatted database and implemented the machine-learning model. The medical members assessed which variables made clinical sense, how they should be interpreted and whether the proposed intervention reflected the realities of patient care.
Hrvoj said the team had participated in hackathons before, although the AI4Health.Cro competition demanded a more evidence-driven approach. Participants worked with a defined healthcare dataset and a specific clinical problem. Every technical decision therefore had to remain connected to its medical relevance.
The clinicians helped the developers distinguish meaningful patterns from correlations that might appear statistically interesting while offering little practical value. Their involvement also influenced the design of the voice assistant and the information presented to doctors.
This form of collaboration is essential in health technology. A technically robust model can still fail when its outputs are poorly aligned with clinical decisions, patient behaviour or the time available to healthcare professionals.
What the competition made possible
AI4Health.Cro gave the Insulogic team access to a clearly defined clinical challenge, a structured healthcare dataset and a working environment in which medical and technical expertise could develop together.
The programme helped the team turn an initial concept into a functioning prototype that combined risk prediction with direct patient follow-up. Working with healthcare data also revealed the demands of the next stage of development, including the need for stronger data science capacity and formal clinical validation.
The competition created a setting in which the team could investigate the problem, test assumptions and examine how a predictive model might become part of a usable healthcare service. Medical guidance was especially important when selecting relevant variables and deciding how the system should respond after identifying a patient at risk.
The award provides funding for further technical development, while the professional contacts established through AI4Health.Cro could support a pilot with a hospital or primary healthcare centre.
The programme also gave the team a clearer picture of the expertise required to take the platform further. That clarity is an important outcome in early-stage health innovation, where the composition of the development team can determine the reliability of the eventual product.
Beyond diabetes
Insulogic emerged from a challenge focused on adherence in type 2 diabetes, although its underlying approach could be relevant to other areas of chronic care.
The same combination of risk detection and automated follow-up could potentially support people taking long-term medication for cardiovascular disease, respiratory conditions or other illnesses requiring sustained treatment.
The team also sees possible applications in monitoring side effects among older patients and following recovery after surgery.
Each of these settings contains a similar gap. Clinicians need information about what happens between appointments, while their capacity for regular personal follow-up remains limited.
A voice assistant could collect routine updates, identify changes and direct professional attention towards patients whose answers suggest that further assessment is needed. Its usefulness would depend on restraint: the system should escalate relevant information without overwhelming clinicians or weakening direct relationships with patients.
Insulogic is still an early-stage prototype. Its predictive accuracy, safety, acceptability and clinical benefit remain to be established through prospective testing. The project nevertheless addresses a central question in healthcare AI. Prediction becomes meaningful when it is connected to a timely response that fits within clinical practice.
AI4Health.Cro helped the team explore that connection using health data, medical expertise and a defined pathway towards further development. The next stage will determine whether the system can make the same transition in practice, from an intelligent model to a dependable part of patient care.
For teams working on data-driven healthcare solutions, the AI4Health.Cro Innovation Challenge provides access to clinical problems, structured health data and interdisciplinary expertise. It creates an opportunity to test ideas against healthcare needs, identify weaknesses early and prepare promising prototypes for validation, piloting and eventual integration into care.