My interest in combining healthcare science with computer science from seeing firsthand how technology supports clinical laboratory processes. Whilst working on a Laboratory Information Management System (LIMS) upgrade, I became increasingly curious about the potential of digital technologies to improve the way healthcare information is processed and used.
This experience helped shape the direction of my MSc Independent Research Project (IRP), where I wanted to explore how artificial intelligence could be applied to a familiar clinical laboratory challenge: identifying potentially anomalous biochemical test results.
Can Machine Learning Improve Laboratory Validation?
Clinical biochemistry laboratories process thousands of patient samples, producing results that inform important clinical decisions. Before these results are released, they must be checked for accuracy and clinical plausibility to support patient safety.
Traditionally, laboratories use rule-based validation systems, which identify potentially abnormal results using predefined criteria such as reference ranges, comparisons with previous patient results (delta checks) and logical consistency rules. However, advances in machine learning have raised questions about whether more adaptive approaches could improve anomaly detection.
To investigate this, I developed a clinically informed synthetic dataset containing 5,000 observations, including six routinely requested biochemical tests, demographic information, previous results and five categories of anomalies. Using synthetic data allowed me to evaluate the methods in a controlled environment without using confidential patient information.
I compared a traditional rule-based validation system against two unsupervised machine learning algorithms, Isolation Forest and Local Outlier Factor. Their performance was assessed using precision, recall, F1-score and anomaly-specific detection rates.
What Did I Discover?
Perhaps the most surprising part of my research was discovering that the traditional rule-based system outperformed both machine learning algorithms. I had initially expected machine learning to perform better, particularly when identifying more complex anomalies, but the results demonstrated the importance of testing assumptions rather than simply expecting newer technologies to be more effective.
The rule-based system achieved the strongest overall performance, correctly identifying 491 of 500 anomalous observations without any false positives and achieving an F1-score of 0.991. In comparison, Isolation Forest identified 340 anomalies (F1-score: 0.680), while Local Outlier Factor identified 163 (F1-score: 0.326).
These findings demonstrated that well-designed rule-based systems remain highly effective at identifying predefined biochemical abnormalities. However, machine learning may still offer value as a complementary approach, particularly for identifying patterns that existing rules may not capture.
As the research used synthetic data, further evaluation using real clinical datasets would be needed before considering implementation in laboratory practice.
Ultimately, the project reinforced the importance of evaluating emerging technologies against established methods, particularly where patient safety is concerned.
Reflecting on My MSc Project
Completing my MSc IRP has been both challenging and rewarding. Transitioning from Biomedical Science into Computer Science meant developing new programming and analytical skills while balancing postgraduate study with full-time employment and family life.
One of the most valuable lessons I learned was that developing a machine learning model is only part of the challenge.
One of the most valuable lessons I learned was that developing a machine learning model is only part of the challenge. Ensuring the dataset is appropriate, clinically meaningful and suitable for evaluating the methods is equally important. My supervisor’s feedback was particularly helpful in encouraging me to think critically about the realism of my synthetic data and the limitations this could introduce.
Although the process involved plenty of problem-solving, seeing the dataset and detection methods come together was incredibly satisfying. It also helped me appreciate the importance of data quality when evaluating machine learning techniques.
During my studies, I also transitioned into working in digital systems within the NHS, supporting Electronic Patient Record (EPR) projects. This has allowed me to continue exploring the relationship between healthcare and technology in a professional setting.
Looking back, one of the most valuable aspects of the project was realising how much my background in Biomedical Science contributed to my understanding of the computational problem. What initially felt like a transition into an entirely different discipline became an opportunity to bring both areas together.
I would also like to thank my research supervisor for their guidance, feedback and support throughout this project.
Leave a Reply