Artificial Intelligence: From Intuition to Evidence

Data-Driven Medicine: The Rise of Machine Learning

Author: Dr Marika Kaden, Mittweida University of Applied Sciences, 19 March 2026

The term Artificial Intelligence (AI), coined by John McCarthy in 1955, has evolved since the 1990s from a largely theoretical research field into a key technology with significant societal and economic relevance. This is particularly evident in areas characterised by large and complex datasets, such as medicine, materials science, climate modelling and language processing. While the term “AI” remains widely used, it is often more accurate to speak of Machine Learning (ML). AI frequently evokes associations with human-like intelligence and autonomy, whereas ML more precisely describes the underlying data-driven algorithms and methods. In many fields, ML is transforming not only technological possibilities but also ways of working – moving from experience-based intuition towards data-driven evidence.

Imagine an internal medicine practice in a medium-sized town in Saxony, such as Mittweida. In addition to a conventional electronic patient record system, the physician uses an ML-based clinical decision support system. This system has been trained on historical patient data, clinical guidelines, laboratory results and imaging findings and is continuously updated with new anonymised data.

A middle-aged patient presents with non-specific symptoms: fatigue, occasional shortness of breath when climbing stairs and diffuse chest pain. Taken individually, these symptoms are unremarkable and could indicate anything from harmless conditions to serious cardiovascular disease. Within seconds, the ML system analyses symptoms, medical history, vital signs, laboratory values and medication records, generating a prioritised list of possible diagnoses together with estimated probabilities and recommendations for further investigations.

Importantly, ML methods do not make diagnoses in the legal or medical sense. Instead, they generate statistical risk profiles that support the physician’s clinical judgement. Final responsibility remains with the clinician, who often possesses additional contextual knowledge about the patient. Nevertheless, these systems can highlight rare or unfamiliar conditions that might otherwise be overlooked.

Modern ML approaches share several key characteristics:

• Data-Driven Learning: Models learn probability distributions from example data rather than relying on explicitly programmed rules.
• Non-Linear Modelling: ML systems can process high-dimensional data and identify complex relationships in images, signals and text.
• Generalisation Rather Than Memorisation: The goal is to extract patterns that can be transferred to previously unseen data.

In practice, different architectures are tailored to specific tasks, including Convolutional Neural Networks (CNNs) for image analysis, Graph Neural Networks for network-based data and Transformer architectures for language and sequence processing. Despite their capabilities, ML systems are not conscious entities; they are mathematical optimisation procedures designed to minimise objective functions.

From a scientific and healthcare system perspective, there are three key opportunities:

1. Improved Diagnostic Precision
   ML systems can detect subtle patterns that are difficult for humans to recognise, such as microscopic abnormalities in histopathology images or temporal patterns in biosignals. For some tasks, they already achieve accuracy levels comparable to or exceeding those of human experts.

2. Scalability and Efficiency
   Many healthcare processes are standardised and data-intensive. ML can support routine tasks and partially automate workflows, helping to reduce workload pressures and allocate resources more effectively. Significant efficiency gains have also been demonstrated in hospital logistics and bed management.

3. Personalised Medicine
   By integrating diverse datasets, including genomics, imaging, laboratory results and clinical histories, ML models can generate personalised prognostic assessments and treatment recommendations. They can identify patient subgroups that respond differently to therapies and thereby support more precise treatment approaches.

ML systems also have important limitations:

• Data Quality and Representativeness
  Models learn only from the data available to them. If datasets are biased or unrepresentative, those biases may be reproduced or amplified.

• Explainability and Transparency
  Many high-performing models function as black boxes. In safety-critical fields such as medicine, there is a legitimate need for transparent and understandable decision-making processes.

• Responsibility and Regulation
  Within the European Union, many healthcare-related ML systems are expected to fall into the category of high-risk AI systems, requiring strict standards for safety, transparency, validation and monitoring. Responsibility remains with healthcare professionals and operating organisations.

• Dependence and Loss of Expertise
  Overreliance on ML systems may erode human expertise if critical skills are increasingly delegated to machines. ML should therefore be understood as an assistive technology and integrated into education and training rather than replacing professional expertise.

Artificial Intelligence – more precisely, Machine Learning – is transforming medicine not through science-fiction scenarios, but through the gradual transformation of routines, decision-making processes and data flows. It offers the opportunity to make diagnostics and treatment more evidence-based, reproducible and personalised. At the same time, it requires high standards in data quality, validation, governance and ethical reflection.

The strength of ML lies in processing large volumes of data; the strength of humans lies in contextual understanding, judgement and responsibility. A scientifically grounded perspective on ML therefore requires the deliberate combination of both capabilities.