The best explanation is a shortlist

A patient arrives with fever, a stiff neck and a headache that started yesterday. Nobody in the room deduces the diagnosis, and nobody waits for enough cases to generalise from. The clinician asks a different question: what would best explain this? Meningitis would. So would a few other things, and the next hour is spent telling them apart.

That is abductive reasoning — inference to the best explanation. Deduction goes from rule to certain conclusion. Induction goes from many cases to a general rule. Abduction goes from an observation back to its most plausible cause, and it is how most of clinical work actually happens: diagnosis, reading a lab result that does not fit, working out why a claim was rejected.

Where AI fits

Modern models are good at the core move of abduction: given a messy set of evidence, propose explanations and rank them. In healthcare software that shows up in a few concrete places.

• Differentials. Symptoms, vitals and results in; a ranked list of candidate conditions out.
• Clinical coding. A discharge summary in; the ICD-10 codes that best explain what was documented out.
• Results that don’t fit. A value that contradicts the rest of the picture; candidate causes, from a real change to a haemolysed sample.

All three share a shape. The input is incomplete, several answers are plausible, and the job is to choose between them.

Where it goes wrong

Abduction has a known weakness, and AI inherits it. The “best” explanation is only the best among the ones considered. A model will always produce an answer, fluently, even when the true cause is not on its list — or when the right answer is “not enough information yet”. A human clinician says “let’s run another test”. A model, left alone, says something confident.

The failure is also quiet. A wrong code or a missed differential looks exactly like a right one on screen. It is found later, by a claims reviewer or a colleague, if it is found at all.

Design rules we hold to

These follow directly from how abduction fails, and they are the rules we design AI assistance toward.

1. Constrain the candidates. Abduction is only as good as the explanations on the table. Give the model a real list to choose from, and it cannot invent an answer that doesn’t exist.
2. Show the shortlist, not just the winner. A ranked list of alternatives keeps the clinician reasoning. A single answer invites them to stop.
3. Show the evidence. A suggestion should point back to the part of the record that supports it, so a human can check the reasoning, not just the result.
4. Let it abstain. “Insufficient documentation” is a valid output. A model that cannot say it will make something up instead.
5. A human confirms. The suggestion is a draft. The coder or clinician who accepts it is the one accountable, and the audit trail records who that was.

The model proposes explanations. A person decides which one is true.

In Pakarti Utama, ICD-10 suggestions start with the first rule. The clinical text is first matched against the hospital’s own ICD-10 catalog by keyword search, so the language model chooses from real candidates rather than generating a code from memory. A validator then rejects any code that doesn’t exist in the catalog, and what survives appears as suggestions in the coder’s code picker, where the coder makes the choice.

That is less impressive than a system that “just codes”. It is also the version that holds up when an auditor asks why a code was chosen, and who chose it.