The problem comes before the technology
AI creates value only when it is grounded in a real healthcare process
Artificial intelligence can help analyse growing volumes of data, anticipate events, reduce manual work and deliver important information faster. A model alone, however, is not a working solution.
It needs appropriately prepared data, a testable research method, integration with existing systems and an interface that users can understand. The boundaries of its use must be explicit, and people must retain control of decisions.
We do not create AI for technology’s sake. We create solutions for the real needs of healthcare organisations.
When does an AI project become R&D?
Research and development starts where material uncertainty exists: the data may not support the expected outcome, standard methods may be insufficient, several approaches need comparison, or an experimental result still has to be turned into a deployable system.
- the required quality cannot be assumed in advance;
- the method needs to be developed or materially modified;
- alternative models and validation methods must be compared;
- the result must work in conditions not reflected by off-the-shelf products.
Six questions before selecting a model
- What problem are we solving?
- What will count as a measurable result?
- Which data is available?
- What remains unknown?
- How will the hypothesis be verified?
- How will the research result be used in practice?
AI and models
We select methods for the problem—not the problem for the method
We compare statistical approaches, classical analytical methods and machine-learning models. Greater complexity must be justified by better quality, stability, interpretability or usefulness in the target process.
Machine learning and predictive models
Demand forecasting, event classification, risk identification, anomaly detection, prioritisation and explainable recommendations.
Time-series analysis
Trends, seasonality, cycles, sudden changes, delayed relationships and differences between healthcare organisations or organisational units.
Event classification and prioritisation
Assessing whether an event requires attention, assigning priority and recipient, and supporting repetition, grouping or controlled escalation.
Natural-language processing
Extracting information from documents, classifying correspondence, identifying obligations and deadlines, analysing contracts and linking messages to the correct case.
Federated learning
Developing shared models while source data remains in participating organisations, with research covering data heterogeneity, update security, versioning and privacy risks.
Explainability and interpretation
Showing influential data, detected risk, model confidence, the period covered by the data, possible alternatives and situations in which a result may be less reliable.
The foundation of every AI project
The hardest problem is often not the algorithm, but the data
Before expensive model training starts, we assess whether the available information can answer the research question credibly.
We assess data quality and meaning
- completeness and timeliness;
- consistency between systems;
- duplicates and outliers;
- changes in recording practices;
- identifiers, units and dictionaries;
- class balance and rare cases;
- systematic error and bias;
- the risk of learning incidental correlations.
If data quality does not support credible research, we say so before the project commits resources to training models. We then identify what should be improved, supplemented or measured first.
From experiment to product
A model must become part of a reliable information flow
An AI component cannot remain an isolated notebook or demonstrator. It needs current data, a suitable user interface, controlled actions and the engineering needed for safe operation and maintenance.
Production-grade integration
- source-system downtime handling;
- queueing and retrying operations;
- data-flow monitoring and error diagnostics;
- access control and event logging;
- interface versioning;
- secure exchange with external systems.
Productisation of research outcomes
- stable data pipelines and a user interface;
- authorisation and exception handling;
- security and acceptance testing;
- model and application version management;
- rollback and controlled release mechanisms;
- post-launch quality monitoring and documentation.
problem→hypothesis→data→experiment→prototype→pilot→implementation→development
Responsible AI
Useful models need clear limits, human oversight and monitoring
We evaluate not only whether a model works on average, but also where it can fail and what the consequence of an error would be in the real process.
People remain in control
The division of responsibility between the system and the user must be explicit. In processes that may affect patient care, final assessment belongs to an authorised professional.
One metric is not enough
We examine sensitivity, specificity, precision, false-positive and false-negative results, performance for individual groups and the impact of errors on work.
Scope of use is defined
When input data differs materially from the research data, the result may require additional caution or rejection rather than automatic use.
Models are monitored
After deployment, we monitor input quality, result distribution, alert frequency, recommendation acceptance, disagreement cases and quality degradation.
Our solutions are intended to support healthcare staff and managers. They do not remove professional responsibility and must not conceal uncertainty behind an apparently definitive recommendation.
AI and R&D in practice
Two projects, two different information challenges
MedAlert addresses information overload and routing in hospitals. Pharma Flow is being developed to support forward-looking medication management in hospital pharmacies.
Status: R&D completed · preparation for implementation
MedAlert
Intelligent selection and routing of information that may require attention from healthcare staff.
- detection and categorisation of events;
- priority and recipient selection;
- message-presentation and escalation paths.
Status: R&D project in progress
Pharma Flow
The planned solution will use AI and federated learning to support hospital-pharmacy medication management.
- it will forecast medicine demand and support inventory optimisation;
- it will identify potential shortages and excessive stock;
- it will support contract monitoring, supplier communication and intervention orders.
Science and engineering
An interdisciplinary team connects research with software delivery
The R&D team works directly with software engineers, healthcare organisations and scientific partners. Here we present current roles and key capabilities; full biographies remain available on the About page.
Aldona Rosner, PhD Eng.
Head of Research and Development
She leads R&D work, ensuring methodological consistency, delivery of the research plan and translation of research outcomes into system prototypes.
Academic profile of Aldona Rosner
Małgorzata Bach, DSc, PhD Eng.
Expert in artificial intelligence, machine learning and data science
She designs and evaluates machine-learning methods and AI models used in the MedAlert and Pharma Flow projects.
Academic profile of Małgorzata Bach
Aleksandra Werner, PhD Eng.
Expert in data, machine learning and research architecture
She combines expertise in databases, data architecture and machine learning with practical applications of AI in medicine.
Academic profile of Aleksandra Werner
Krzysztof Werner, MSc Eng.
Researcher and software engineer
He combines research with prototyping, technical validation and the development of working software components.
Academic profile of Krzysztof WernerA verifiable delivery path
Nine stages from the problem to monitored use
The process remains iterative, but each stage has a clear purpose and provides evidence for the next decision.
- 01
Understand the problem
Talk to future users and examine the real process.
- 02
Assess the data
Check its meaning, quality, completeness and lawful availability.
- 03
Formulate hypotheses
Define uncertainty and measurable evaluation criteria.
- 04
Design experiments
Select methods, datasets and a validation approach.
- 05
Compare alternatives
Test several approaches rather than one preselected algorithm.
- 06
Build a prototype
Connect the model to an interface and a representative workflow.
- 07
Validate with users
Evaluate the result, clarity, ergonomics and effect on work.
- 08
Integrate and deploy
Turn research outcomes into a secure, maintainable solution.
- 09
Monitor
Control data quality, model behaviour and actual use of recommendations.
Cooperation models
We can deliver the whole path or a defined part of it
The cooperation model depends on the maturity of the idea, available data, partners and the target implementation environment.
End-to-end AI project
From concept and data assessment through a prototype, integration and implementation.
Partner in an R&D project
Participation in consortia with healthcare organisations, universities and companies.
Selected research or engineering task
Data preparation, AI models, an experimental environment, prototype or integration.
Idea and data audit
Assessment of whether AI is justified and whether credible research can begin.
From prototype to product
Transforming a research model or demonstrator into a deployable solution.
EU-funded projects
From the research concept and technology section of the application to implementation of results.
Let us start with the problem
You do not need to arrive with a ready-made AI model specification
A better starting point may be a manual analytical process, a risk detected too late, dispersed data, delayed information or a research prototype that has not reached implementation.
Together, we will assess whether the right response is AI, classical analytics, process automation, systems integration—or a combination of these approaches.