Decoded #4: Pharmacist in the Loop

Human-, clinician-, or pharmacist-in-the-loop have taken over health tech AI pitch decks; however, to each company it means something different. The key differentiator is where the healthcare professional (HCP) enters the process.

Step 1: Translating Clinical Reasoning to Machine Logic (AKA "The Engineering")

Clinical judgment is very abstract and views the world in gray, constantly weighing benefit/risk trade-offs in every case. Machines, on the other hand, thrive on black-and-white patterns and repetition. Therefore, a universal language between the two must be created, which is also why a well-structured deterministic system is the most consistent approach in present AI models. An example of this may be an HCP assigning a basic algorithm or blueprint, but then including rigid rules/variables to continue to keep it simple for the machine to understand yet keep the tool grounded in evidence. (e.g. Metoprolol succinate = good; tartrate = bad in HFrEF)

Step 2: Coding and Prompt Generation of the Decision Engine (AKA "The Assembly")

This is where the model comes to life and the architecture is built line by line for the machine to "reason." Coded machine language in a JSON file is a common coding prompt, but for HCPs this is a foreign dialect and, quite frankly, intimidating. However, this is where a seasoned developer or AI tool (e.g. Claude Code) can help translate that JSON file into clinician language. This allows clinicians to get under the hood and audit the code in depth to identify misinterpretation or "teach" the software nuances that only years in the field can offer.

Step 3: Running Test Scenarios/Auditing Output (AKA "The Test Driver")

This strategy is the most straightforward and where HCPs try to drive and "break the tool." This is practice mode, trying to find weaknesses or blind spots where incorrect answers are in the output and a plan is developed to correct the issue.

Some combination of steps 1 and 3 is where most health tech companies live; however, we believe the team that leaves no stone unturned in all steps will be the most sound. Think of it like this: if only doing steps 1 and 3, the HCP knows the design and how to drive it, but at some point the tool will inevitably "break" or reason improperly (e.g. guideline change that has not been adapted into the reasoning). The team that had the HCP involved in assembling every bolt, cylinder, and weld will know exactly how to correct and perfect the output at the source. This adds a safety layer that allows the HCP to be the direct translator for the machine to ensure the same mistake never happens twice.