Why MYCIN Used Certainty Factors Over Bayesian Probabilities

In the 1970s, the groundbreaking artificial intelligence expert system MYCIN was developed at Stanford University to diagnose bacterial blood infections and recommend antibiotic treatments. Despite the established mathematical rigor of Bayesian probability for handling uncertainty, MYCIN's creators opted to invent a new method called Certainty Factors (CF). This article explores the primary reasons behind this decision, including the lack of comprehensive medical data required for Bayesian models, the invalid assumption of conditional independence among symptoms, and the fundamental disconnect between how human doctors reason and how pure probability theory operates.

The primary obstacle to using pure Bayesian probability in MYCIN was the massive amount of statistical data it required. To apply Bayes' theorem accurately, the system needed to know the prior probabilities of all possible bacterial infections and the conditional probabilities of every single symptom given each infection. In the 1970s, and even today, this granular level of clinical data simply did not exist. Medical literature rarely recorded the exact numerical probability of a specific symptom occurring in the presence of a rare pathogen, making a strictly Bayesian approach impossible to populate with accurate data.

To work around the lack of data, Bayesian systems often rely on the assumption of conditional independence—the idea that symptoms occur independently of one another given a specific disease. In medicine, this assumption is frequently false. Symptoms are often biologically linked and interdependent; for example, a high fever and an elevated white blood cell count are both parts of the body's inflammatory response. Ignoring these dependencies in a Bayesian network leads to distorted and dangerously inaccurate diagnostic probabilities.

Furthermore, Bayesian probability requires clinicians to adhere to a strict mathematical rule: the probabilities of an event happening and not happening must always sum to one. When the developers of MYCIN interviewed infectious disease experts, they discovered that human doctors do not think this way. If a symptom strongly suggests a specific bacterium, a doctor's belief in that diagnosis increases. However, the absence of that same symptom does not necessarily mean the bacterium is not present. Pure probability forces a direct link between belief and disbelief, whereas medical experts weigh evidence FOR a diagnosis separately from evidence AGAINST it.

To align the AI with human cognitive processes, Edward Shortliffe and Bruce Buchanan developed Certainty Factors. This calculus represented uncertainty on a simple scale from -1.0 (definitely false) to +1.0 (definitely true), with 0 representing complete ignorance. Crucially, the system maintained two separate values for every hypothesis: a Measure of Belief (MB) and a Measure of Disbelief (MD). This allowed MYCIN to gather positive and negative evidence independently, combining them into a final Certainty Factor only after all rules were evaluated.

Finally, MYCIN was built as a modular, rule-based system using ratusan of "IF-THEN" statements. In a pure Bayesian system, adding a single new medical rule or symptom would require recalculating the conditional probabilities across the entire database to maintain mathematical consistency. Certainty Factors allowed developers to add, remove, or modify individual rules in isolation. Because each rule carried its own self-contained Certainty Factor provided by a human expert, the system could easily grow and adapt without breaking the underlying mathematical framework.