Three terms, one persistent confusion

“Artificial intelligence,” “machine learning,” “deep learning,” “algorithm,” “neural network”… These terms circulate in medical conferences, in the specialist press and in the sales pitches of healthcare software publishers. They are often used interchangeably, sometimes by the vendors themselves, sometimes by ill-informed journalists, and sometimes even by AI professionals who oversimplify.

The result: many healthcare professionals have the impression of vaguely understanding what it is all about, without being able to really distinguish these concepts or concretely evaluate what a given tool actually does.

This confusion has practical consequences. A doctor who does not understand the difference between a rule-based expert system and a deep learning model cannot correctly evaluate the limits of the tool being offered to them. They cannot ask the right questions of a manufacturer, nor understand why an algorithm can perform brilliantly on one population and fail on another.

This article offers you a progressive clarification anchored in concrete medical examples, without mathematical formulas, without unnecessary computing jargon.

The Russian dolls image: a metaphor to understand it all

Imagine three Russian dolls nested one inside the other. The largest, which contains all the others, is called artificial intelligence. Inside is a smaller doll called machine learning. And inside that one, an even smaller doll: deep learning.

In other words: all deep learning is machine learning, and all machine learning is AI. But the reverse is not true: not everything that falls under AI is machine learning, and not everything that falls under machine learning is deep learning.

Artificial intelligence: the general framework

Definition

Artificial intelligence is the field of computer science that aims to design systems capable of performing tasks which, if carried out by a human being, would require intelligence: understanding a text, recognising an image, solving a problem, making a decision, translating a language.

The “classic” approaches to AI in healthcare

Before machine learning and deep learning, AI in medicine relied mainly on expert systems, computer programs encoding logical rules formulated by human experts of the type “IF the patient has a fever AND a productive cough AND an opacity on the X-ray THEN consider bacterial pneumonia.”

These expert systems had their heyday in the 1980s-1990s. MYCIN for bacterial infections, INTERNIST-1 for internal diagnosis, DXplain for differential diagnosis. They have the merit of being transparent and explainable; one can know exactly why the system produced a given recommendation. But they have a fundamental limitation: they can only incorporate the rules that a human expert was able to formulate explicitly, and they quickly prove insufficient in the face of the complexity and variability of real clinical situations.

Machine learning: learning from data

The paradigm shift

The fundamental difference between a classic expert system and a machine learning system comes down to one sentence: the expert system follows rules written by humans, the machine learning system discovers rules on its own from data.

Instead of explicitly programming “if A and B then C,” the system is given thousands of examples of A, B and C, and is left to identify on its own the patterns that make it possible to predict C from A and B. This is what is called machine learning, or training the model.

How does it work in concrete terms?

Let us take a simple medical example. You want to train a machine learning model to predict the risk of readmission within 30 days after a cardiac decompensation.

You assemble a historical dataset of 50,000 hospital stays for heart failure, annotating for each whether the patient was readmitted within 30 days or not. Each stay is described by a set of variables: age, sex, ejection fraction, BNP level, serum creatinine, number of medications, length of stay, history…

The machine learning algorithm will analyse these 50,000 examples and identify the combinations of variables that are statistically associated with readmission. It will produce a mathematical model which, faced with a new patient at discharge, will calculate an individual risk score.

The main families of machine learning algorithms used in healthcare

Logistic regression is the most classic and most transparent approach: it produces a linear equation linking input variables to a probability of outcome. The Framingham score for cardiovascular risk is a familiar example.

Decision trees and random forests build tree structures of successive rules. Random forests are widely used in predictive medicine for their good performance/interpretability ratio.

Support vector machines (SVM) are effective for classification problems with little data and many variables, a profile frequent in medical genomics.

Gradient boosting (XGBoost, LightGBM) is today one of the most high-performing algorithms on structured tabular data, exactly the type of data contained in patient records.

The limit of classic machine learning

These algorithms excel on structured, tabular data. But they prove insufficient in the face of unstructured data of high complexity: a medical image containing millions of pixels, a continuous ECG signal, a report written in free language. It is to process these types of data that deep learning emerged.

Deep learning: the power of deep neural networks

The biological inspiration

Deep learning is inspired, in a very schematic way, by the functioning of the human brain. It uses artificial neural networks: mathematical structures composed of layers of interconnected nodes, each node processing a piece of information and passing it on to the following nodes with an adjustable weighting.

The adjective “deep” refers to the number of successive layers in these networks; modern architectures can have hundreds, even thousands of them.

What deep learning makes possible in medicine

In medical imaging, deep learning has brought about a revolution. A convolutional neural network (CNN) can analyse a medical image pixel by pixel, identify anatomical structures, detect anomalies and locate them with remarkable precision.

In medical language processing, deep learning architectures called transformers (of which GPT and BERT are the best-known representatives) make it possible to analyse and generate medical text with unprecedented sophistication. It is this technology that lies at the heart of medical transcription assistants, automatic report-coding systems, and medical information chatbots.

In biological signal analysis, deep learning makes it possible to analyse continuous ECGs to detect arrhythmias, to interpret electroencephalogram (EEG) traces to identify subclinical epileptic seizures, or to analyse respiratory sounds to detect a bronchospasm.

The downside: the “black box”

The power of deep learning comes with a major problem for medical use: explainability. Unlike a decision tree or a logistic regression whose rules can be read, a deep neural network is a “black box”; it produces a result without it being easy to understand why.

This is why an entire field of research has developed around the explainability of AI (XAI — Explainable Artificial Intelligence), with techniques such as heatmaps that colour the areas of the image that contributed most to the algorithm’s decision.

Recap: what distinguishes these three levels

Level	Clinical analogy	Strengths	Limits
AI (general framework)	Medicine as a whole	Vast field, multiple approaches	Term too broad to be precise
Machine learning	Clinical epidemiology	Tabular data, predictive scores, interpretability	Ineffective on images, text, signals
Deep learning	The highly trained visual expert	Images, text, complex signals	“Black box,” needs massive data

Why this distinction matters for your practice

Evaluating the relevance of a tool. A machine learning algorithm trained on data from Caucasian patients in a Paris university hospital may perform differently on your patients, depending on your context. Knowing that it is machine learning, and therefore a statistical model dependent on its training data, helps you ask the right questions about its external validity.

Interpreting the results with the appropriate perspective. A risk score produced by a gradient boosting model is a statistical probability, not a diagnosis. A deep learning result in imaging is a suggestion to be integrated into your overall clinical reasoning, not an absolute truth.

Identifying limits according to the type of data. If a classic machine learning tool is used to analyse medical images, be wary: this type of algorithm is not architecturally suited to this type of data. A credible tool in medical imaging will use deep learning.

Dialoguing with manufacturers and IT teams. When a publisher presents you with a solution, knowing how to distinguish an expert system from a machine learning model and a neural network allows you to ask precise questions: on what data was it trained? What is its performance in my population? How does it handle atypical cases? Can it explain its decisions to me?

Three concepts, one same requirement of critical scrutiny

Artificial intelligence, machine learning, deep learning: these terms do not designate mysterious technologies reserved for engineers. They are mathematical tools with strengths, limits and specific contexts of use, exactly like the clinical tools you have mastered.

A CT scan is not an MRI, and a blood test is not a biopsy: each tool has its area of excellence and its blind spots. The same applies to the different types of AI. Understanding them at this level allows you to use them with discernment, to get the most out of them and to protect your patients from their limits.

And this understanding, contrary to what one might think, does not require advanced mathematical training. It simply requires taking the time to demystify, which is precisely what you have just done by reading this article.

Would you like to deepen your understanding of AI and learn how to concretely evaluate the tools offered in your specialty? Our training courses are designed for healthcare professionals, with no technical prerequisites. Discover the programme.

Sources and references

1. Shortliffe, E. H. & Buchanan, B. G. (1975). “A model of inexact reasoning in medicine.” Mathematical Biosciences, 23(3-4), 351–379. [MYCIN system]
2. Miller, R. A. et al. (1982). “INTERNIST-1, an experimental computer-based diagnostic consultant for general internal medicine.” New England Journal of Medicine, 307(8), 468–476.
3. Barnett, G. O. et al. (1987). “DXplain — An evolving diagnostic decision-support system.” JAMA, 258(1), 67–74.
4. LeCun, Y., Bengio, Y. & Hinton, G. (2015). “Deep learning.” Nature, 521(7553), 436–444.
5. Rajkomar, A., Dean, J. & Kohane, I. (2019). “Machine learning in medicine.” New England Journal of Medicine, 380(14), 1347–1358.
6. Vaswani, A. et al. (2017). “Attention is all you need.” Advances in Neural Information Processing Systems, 30. [Transformer architecture behind GPT and BERT]
7. Esteva, A. et al. (2017). “Dermatologist-level classification of skin cancer with deep neural networks.” Nature, 542(7639), 115–118.
8. Chen, T. & Guestrin, C. (2016). “XGBoost: A scalable tree boosting system.” Proceedings of the 22nd ACM SIGKDD, 785–794.
9. Ribeiro, M. T. et al. (2016). “‘Why should I trust you?’: Explaining the predictions of any classifier.” Proceedings of the 22nd ACM SIGKDD. [Explainability / XAI]
10. Topol, E. J. (2019). “High-performance medicine: the convergence of human and artificial intelligence.” Nature Medicine, 25, 44–56.