Leveraging AI to build a model for predicting mortality in the Intensive Care Unit, optimizing patient care strategies

Kajaani Central Hospital (KAKS), part of Kainuu wellbeing services county, identified an opportunity to enhance critical care decision-making processes within its Intensive Care Unit (ICU).  The chief physician, who is responsible for making daily critical decisions for critically ill patients, recognized a pressing need: how to harness existing patient data more effectively to support timely and informed clinical decisions.

The ICU generates large volumes of complex patient data daily, including continuous monitoring of vital signs, laboratory results, detailed medical histories, and treatment records. However, much of this data remains underutilized due to the challenge of rapidly processing and interpreting vast volumes of information. The manual review of these complex datasets is time-consuming and place additional demands on medical staff. 

This highlights a vital need for advanced solutions capable of efficiently processing and analyzing ICU data, delivering actionable insights that can be integrated into clinical workflows in real time.

A core challenge emerged:

How to develop and implement a predictive AI model that can support ICU staff by identifying patients at high risk of mortality, based on existing physiological and clinical data.

Public investment is vital for broad access to digital health solutions. Through HealthHub Finland EDIH, which provides free services to public sector entities, Kajaani University of Applied Sciences (KAMK) and the AIKA Ecosystem offered the opportunity for Kajaani Central Hospital (KAKS) and Kainuu Wellbeing Services County to collaboratively co-develop an AI model aimed at improving ICU decision-making and patient care, enabling ethical, secure, and cost-effective AI adoption in regional healthcare.

Key features of Machine Learning (ML) model:

The central objective was to develop and integrate an ML model capable of:

• Analyzing existing ICU patient data in real time

• Identifying patients at high risk of mortality

• Providing actionable insights to assist clinical teams in making faster, evidence-based decisions

Solution Approach

1. Data Utilization and Integration

• Collected and structured historical ICU data.

• Utilised KAKS datasets to train and validate clinically relevant predictive models.

• Recommended additional sources, such as physionet available ICU data, to further enhance model performance.

2. Machine Learning Model Development

• Used advanced ML algorithms to analyse complex patterns and correlations that are often too complex or time-consuming for manual interpretation.

• Allowed the model to predict ICU mortality risk based on a patient’s physiological and clinical data available upon admission and during their stay.

3. Ethics and Patient Privacy

• Established data usage to comply with strict ethical guidelines and patient consent protocols, ensuring responsible and secure handling of sensitive information.

• Prioritised transparency and explainability of AI outputs to maintain clinical trust.

4. Testing and Validation

• Provided initial studies to demonstrate the model’s ability to accurately classify patients by mortality risk, implying that ICU care prioritization and resource allocation could be improved.

• Organised ongoing medical staff feedback to help improve the model´s accuracy over time.

Following a phase of rigorous testing and refinement, the machine learning model developed for predicting ICU mortality has yielded concrete, measurable outcomes that demonstrate its value to healthcare providers and HealthHub Finland EDIH.

The model was initially trained using a dataset of 4,000 anonymized patient records from PhysioNet, incorporating approximately 30 clinical parameters per patient. To evaluate its real-world performance, it was subsequently tested on data from 600 ICU patients at Kajaani Central Hospital (KAKS), with the primary objective of assessing its ability to accurately predict mortality risk.

Implementation testing has shown that the model can help identify high-risk patients in the ICU at an earlier stage, which could facilitate more timely clinical interventions. These capabilities may contribute to reducing the length of ICU stays, optimizing the use of ICU beds, lowering operational costs, and easing the burden on critical care staff.

In partnership with HealthHub Finland EDIH, Kainuu Wellbeing Services County has leveraged this initiative as a step toward strengthening its digital capacity. The project has helped establish the foundations for integrating advanced data analytics and secure data-sharing solutions into the ICU environment. While full-scale deployment is still pending, the work already undertaken has supported the development of the necessary digital infrastructure and expertise to enable future integration of AI tools into everyday clinical workflows.

Funding and Return on Investment

The successful development and delivery of this digital health solution were enabled by a well-coordinated public funding model. HealthHub Finland EDIH’s contribution was made possible through blended funding—50% from the European Commission and 30% from Business Finland. This structure significantly lowered the financial barriers for regional stakeholders and allowed them to focus on innovation, testing, and validation without the pressure of early-stage investment risks.

By reducing upfront costs and accelerating the path from concept to practical testing, this funding approach has laid the groundwork for high-impact returns. These include improvements in operational efficiency, enhanced patient outcomes, and long-term sustainability through the adoption of data-driven, energy-efficient clinical practices. Moreover, the project demonstrates a scalable, ethical, and cost-effective model for implementing AI in healthcare, particularly for publicly funded institutions with limited resources.

Do’s:

• Verify Assumptions Early: Regularly confirm assumptions with clinical staff. This significantly reduces rework and improves the accuracy of data interpretation.

• Understand Medical Context: Medical terms often appear multiple times or in ambiguous ways. Contextual understanding is essential to avoid misinterpretation—especially when terms resemble diagnoses.

• Develop Medical Literacy: Gaining familiarity with medical terminology and scoring systems (e.g., SAPS III) enables more meaningful insights. Investing time in medical reading proved invaluable.

• Maintain File Integrity: Always preserve original data files and use duplicates for processing. This protects data integrity and allows easy backtracking when needed.

• Iterate with Feedback: Presenting findings to the doctors for iterative feedback helps refine results and align technical insights with clinical relevance.

Don’ts:

• Don’t Assume Raw Data is Self-Explanatory: Directly analyzing unstructured medical data without clinical input can lead to incorrect conclusions.

• Don’t Skip Documentation: Failing to document steps and reasoning slows collaboration and increases the chance of repeated mistakes.

• Don’t Expect Perfection: Mistakes and refinement are part of the process. They should be seen as growth opportunities rather than failures.

Recommendations for Future Use:

• Involve domain experts from the start to avoid data misinterpretation.

• Plan for dedicated time to build up domain knowledge.

• Create clear file management protocols to avoid data loss or confusion.

• Build a feedback loop with stakeholders to continuously validate work.

These lessons offer practical advice for EDIHs and SMEs/PSOs engaging in data-driven healthcare projects. Future projects will benefit from these structured practices and collaborative approaches.