Component
AI integrated rule discovery
Finding the simplest model that explains the data, and deliberately avoiding black-box complexification — because a decision affecting a patient has to be one you can explain to them.
Why this exists
There is an obvious temptation in a program with this much data attached to it: build a large model, let it find patterns nobody specified, and act on its output. IntellaRx does not do that, and the reason is not technical modesty.
It is that this program exists because of a system that made consequential decisions about people using an instrument whose logic nobody had to justify. Replacing an opaque count-based rule with an opaque learned rule would reproduce the exact failure described in step six with better mathematics and worse accountability.
A patient is entitled to know why something happened to them. A clinician is entitled to disagree with a system and be able to say what they are disagreeing with. An agency is entitled to have its analysts audit a rule. None of those survives a model that cannot be inspected.
How it works
- Simplest model that fits. Where a small set of interpretable rules performs comparably to a complex model, the interpretable rules win, and the comparison is reported rather than asserted.
- Rules are stated in clinical language, so they can be argued with by the people they affect.
- Discovery, not decision. The output is a candidate rule for clinical review. Clinicians decide whether it makes sense and whether it should be used at all.
- Performance is reported with its failure modes — including which subgroups it does worse on, which is where automated systems in health care usually go wrong.
- Rules are versioned and dated, so that any decision can be traced to the rule in force at the time.
What this is not
It is not a scoring system that ranks patients by suspicion. The program does not build one, and an agency asking for one is asking for something outside this scope.
It is not automated decision-making about an individual’s treatment. A rule can prompt a clinician. It cannot act.
It is not predictive policing applied to medicine. Whatever a model may correlate with, the outputs here are used to improve care and to detect changes in a drug supply, not to identify people as future offenders.
It is not trained to identify likely misuse from demographic or social characteristics. Given what the social determinants evidence shows, a model built that way would efficiently reproduce existing disadvantage and call it a finding.
Questions this raises
Why avoid more powerful models if they perform better?
Because performance is not the only requirement. A rule that affects a person’s access to treatment has to be explainable to that person, disputable by their clinician and auditable by an agency. Where an interpretable model performs comparably we use it, and where it does not we report the gap rather than quietly taking the opaque option. See the simplest model that works.
Does IntellaRx build risk scores for individual patients?
No. The program does not produce a per-person suspicion score, and would decline a request to build one. Rules exist to surface clinical considerations and to detect changes in a local drug supply. See what we do and do not do.