Imagine going to your doctor with an infection. But infections can be tricky – different bacteria require different antibiotics and dosages to treat. Now imagine if your doctor had an assistant – an AI system that could recommend what infection you likely have and the best way to treat it.
Back in the 1970s when infectious disease experts were scarce, Stanford researchers built just such a system – MYCIN, the first pioneering expert system applied to medicine.
What are Expert Systems?
First, what exactly are expert systems?
Expert systems are AI programs designed to replicate and automate the decision-making and problem-solving ability of human experts in a specialized domain. They contain a knowledge base of facts, rules, and heuristics that allow them to reason through problems much like a specialist would.
Some advantages of expert systems:
- Can encode rare expertise
- Offer standardized diagnostic capability
- Provide a training mechanism for new doctors
- Offer fast answers without needing an expert on-site
In the 1970s, researchers saw opportunities to use this emerging AI technology in the medical field. And out of this goal, MYCIN was born – an expert system for diagnosing infections.
1970s Medical Landscape – Ripe for Innovation
In the 1970s, while medical technologies were advancing rapidly, there remained challenges in healthcare delivery and expertise shortages. Stanford researchers building MYCIN identified two main issues:
- A scarcity of infectious disease experts capable of identifying unusual infections
- Many general practitioners lacking the specialized skills to diagnose unfamiliar infections themselves
This meant patients often went misdiagnosed or endured lengthy, expensive hospital stays while the rare experts were brought in.
Additionally, identifying not just infection causes but optimal treatments raised further difficulties. Determining which antibiotics in which dosages can require weighing many interdependent factors in real-time.
Automating this specialized diagnostic decision-making is where MYCIN broke new ground.
Introducing MYCIN: Breakthrough Expert System
The MYCIN system was developed over 5-6 years in the 1970s by Edward Shortliffe alongside Stanford clinicians. The initial focus was on blood infections and meningitis diagnoses.
The goal of MYCIN was for its recommendations to match the performance of the best infectious disease experts. To accomplish this, the developers identified two key components needed:
- A comprehensive set of rules encoding infection specialists‘ decision-making methodology
- A system to reason accurately through the rules matching symptoms to likely diagnoses
Over years spent interviewing specialists, the MYCIN team encoded over 600 rules covering infection diagnosis processes into the system. These took the structure of weighted IF-THEN statements, like:
IF
- The stain of the organism is gramneg
- The morphology of the organism is rod
THEN
- The probability the identity is ECOLI is 0.8 (80%)With this knowledge base encoding the nuances of specialist reasoning, the next step was applying it accurately. MYCIN utilized backward-chaining – starting from a set of possible diagnoses and working backward by asking questions to narrow down and confirm which infection was most likely.
Here is a simplified view of how MYCIN walked through this process:
- Take reported symptoms and vital signs
- List possible bacterial organisms based on findings
- Rank organisms by probability based on rules
- Ask diagnostic Yes/No questions to refine probabilities
- Repeat process until target confidence threshold met
- Output: Most likely infection + treatment recommendations
This multi-step diagnostic workflow allowed MYCIN to "think" through infections much as a specialist would.
Groundbreaking Performance
In one early evaluation, MYCIN was tested against 10 actual patient cases involving life-threatening blood infections.
- The gold standards were the treatment plans of two infectious disease experts brought in on each case.
- MYCIN was scored on whether its diagnosis and treatment plan matched the experts.
Results? MYCIN matched or exceeded the experts‘ answers in 8 out of 10 test cases.
Researchers found that MYCIN could accurately diagnose infections and determine optimal antibiotic selection and dosages at an expertise level comparable to the top infectious disease specialists.
This was groundbreaking – an expert system competently mimicking a highly skilled human specialist for the first time.
Advantages and Disadvantages
Some key advantages of MYCIN included:
- Matched or exceeded human specialist performance
- Encoded rare medical expertise allowing standardization
- Could accurately prescribe antibiotics and dosages
- Showed potential of AI decision-support in medicine
However, the system did have some disadvantages:
- Never saw widespread clinical adoption
- Concerns around accuracy outside lab settings
- Rule-based approach made adapting new knowledge difficult
Additionally, computing power in the 1970s imposed technological constraints on the system‘s capabilities versus more modern AI techniques.
Still, while imperfect, MYCIN helped show expert systems could play a real diagnostic support role even with relatively simple software.
The Impact and Legacy
The pioneering MYCIN system demonstrated the possibilities of expert systems and AI in medicine. It provided a template showing that encoding specialist knowledge for computer reasoning could work.
MYCIN helped drive progress on integrating decision-support systems at medical facilities in the decades since. Elements of its structure got incorporated into later systems like DXplain for differential diagnosis and QMR for clinical decision support.
While rule-based systems have downsides, the core ideas behind MYCIN – digitizing human diagnostic processes as computer-executable logic – turned out to be foundationally sound. MYCIN‘s DNA can be seen in many modern medical AI tools playing ever-greater physician-assist roles today.
So while it never became a clinical staple itself, MYCIN helped originate the expert system paradigm that continued advancing toward increasingly powerful modern AI diagnosis.
