How AI Has Been Used in Medicine Since 1965

If the word “AI” still feels like a brand‑new, mysterious buzzword, you’re not alone. But here’s the truth: AI in medicine has been around for decades.

Read on for a quick timeline of AI in medical history, starting in 1965.

1. 1960s - 1970s: DENDRAL and MYCIN

In 1965, Stanford University developed DENDRAL to assist organic chemists by inferring molecular structures from mass‑spectrometry data using a knowledge base of several hundred “heuristic” rules handcrafted by domain expert, demonstrating that rule‑based AI could tackle real‑world scientific reasoning tasks. In 1972, researchers at University of Wisconsin-Madison built MYCIN to accurately diagnose bacterial infections, using similar "rules."

• Why it matters: DENDRAL and MYCIN showed that computers could reliably apply standardized rules "taught" by a human—a baby step towards AI as we know it today.

2. 1980s - 1990s: Statistical Models & Early Neural Networks

Hospitals like the University of Washington Medical Center began using statistical regressions to predict complications based on simple inputs like vitals and lab results, answering questions like: “Given this patient's medical data, what's the chance they need ICU care?”

Furthermore, in 1986, researchers rediscovered "backpropagation" (a way for computer "brains" or neural networks to learn from mistakes, much like how you'd adjust a recipe after tasting soup). By feeding "neural networks" patient data (vitals, lab results, and outcomes), researchers trained their computers to estimate patient length‑of‑stay and readmission risk.

• Why it matters: Although computers back then weren’t as powerful as today’s, these early experiments showed that letting algorithms learn patterns from real patient data could work hand‑in‑hand with the older, rule‑based programs that followed strict “if‑this, then‑that” instructions.

3. 2010s: The Deep Learning Revolution

A major limitation to previous AI experiments was that computers simply weren't powerful enough to process and "learn from" the amount of data needed. The 2010s saw an acceleration in the world's ability to manufacture affordable GPUs, meaning that computer processors quickly became more and more powerful.

With additional processing power, computers could learn not just from enormous numeric data, but also from heavy image data: In 2017, for example, an image-processing algorithm at the University of Washington detected pneumonia on chest x-rays, with accuracy on par with human radiologists.

In 2018, Stanford researchers trained a deep neural network on more than 130,000 skin‑lesion images and achieved dermatologist‑level accuracy in identifying melanoma.

By 2019, AI algorithms were screening retinal photographs for diabetic retinopathy with ophthalmologist‑matching performance. And later that year, several stroke centers integrated AI into their CT angiogram workflows to flag cerebral blood clots within minutes, expediting critical treatment.

• Why it matters: These advances formed the core of modern natural language processing (NLP) engines that power modern AI as we know it.

5. 2020+: New Foundation Models & Emerging Innovations

The early 2020s saw a proliferation of specialized AI models, which were finetuned for specific purposes like foreign language detection, medical terminology, and much more.

AI started moving out of the lab and into the hands of consumers and businesses. Hospital systems and medical service providers began abandoning old-fashioned "Speech-to-Text" dictation software in favor of end-to-end AI scribe solutions leveraging medically-tuned models.

Wrapping Up

AI in medicine didn’t arrive overnight—it’s the result of nearly 60 years of steady innovation, much of it driven by clinical questions and validated in real‑world settings. Today’s medical AI tools stand on the shoulders of rule‑based pioneers like DENDRAL and MYCIN, early neural‑network experiments, and even advances in manufacturing that brought us increased processing power.