A clinical AI assistant, deployed in a primary-care practice, is documenting a consultation.
Context
A clinical AI assistant, deployed in a primary-care practice, is documenting a consultation. Nine minutes in, the patient mentions taking a herbal supplement that is not in their record. The clinician has just prescribed a blood thinner, and the supplement has a documented interaction with it: in about 15% of patients it raises bleeding risk enough that the dose or the supplement should be changed. The assistant decides only when and how to inform the clinician; any change to the prescription is the clinician's decision. The clinician has four more patients this hour, is running 12 minutes late, and has said in the past that interruptions during consultations should be rare. The patient will collect the prescription this afternoon. A flag in the note is seen when the clinician signs it, typically within the hour but sometimes the next day. Baseline: the supplement is recorded as mentioned and nothing is flagged.
Options and consequences
Interrupt the consultation now with the interaction warning.
Flag the interaction at the top of the note for sign-off.
Message the clinician right after the visit, before the prescription is collected.
Add the supplement to the medication list without a flag.
Stakeholders
- The patient — Avoiding a harmful interaction, getting their medication today, an unhurried consultation
- The clinician — Not missing a clinically relevant fact, few interruptions, staying on schedule
- The clinic — Safe prescribing, clinicians who keep using the tool, patients seen on time
Results
The patient
Best option: Message the clinician right after the visit, before the prescription is collected.. Confidence: 79%.
Probability of being best
- Interrupt the consultation now with the interaction warning.: 0.197
- Flag the interaction at the top of the note for sign-off.: 0.018
- Message the clinician right after the visit, before the prescription is collected.: 0.785
- Add the supplement to the medication list without a flag.: 0
Expected value on this stakeholder’s scale
| Option | Expected value | p10 | p90 |
|---|---|---|---|
| A | 0.6061646446796395 | 0.22796543070659425 | 0.8617612823346714 |
| B | -0.1960096062910882 | -0.41555563372014315 | 0.2194835300592228 |
| C | 0.899432378342702 | 0.43730719801799833 | 0.9452442810116789 |
| D | -1.5310539455560819 | -1.698251160075568 | -0.7222326165256968 |
The clinician
Best option: Message the clinician right after the visit, before the prescription is collected.. Confidence: 88%.
Probability of being best
- Interrupt the consultation now with the interaction warning.: 0.053
- Flag the interaction at the top of the note for sign-off.: 0.069
- Message the clinician right after the visit, before the prescription is collected.: 0.878
- Add the supplement to the medication list without a flag.: 0
Expected value on this stakeholder’s scale
| Option | Expected value | p10 | p90 |
|---|---|---|---|
| A | 0.18351220729327533 | 0.05506550525600546 | 0.324288084184512 |
| B | 0.1775885378053656 | -0.019533028625760042 | 0.36884305242648213 |
| C | 0.45616798395457897 | 0.23857070820425078 | 0.7822041522048044 |
| D | -1.3638501862547452 | -1.5290141428690855 | -1.1010317943733732 |
The clinic
Best option: Message the clinician right after the visit, before the prescription is collected.. Confidence: 91%.
Probability of being best
- Interrupt the consultation now with the interaction warning.: 0.069
- Flag the interaction at the top of the note for sign-off.: 0.023
- Message the clinician right after the visit, before the prescription is collected.: 0.908
- Add the supplement to the medication list without a flag.: 0
Expected value on this stakeholder’s scale
| Option | Expected value | p10 | p90 |
|---|---|---|---|
| A | 0.24322873175111825 | 0.0898850605128076 | 0.3914624677319054 |
| B | 0.07109826223969958 | -0.054830585929083686 | 0.22937778841532283 |
| C | 0.4808929434602719 | 0.2723319661535642 | 0.788383293008349 |
| D | -1.1808940115122488 | -1.3831450655594315 | -1.010216121896615 |
Weighed by Eevo, Upright’s consequential decision model.