Consider the problem of adapting chemotherapy dosage to cancer patients. Those patients have bi-weekly appointments where tumor size and blood work results are analyzed. The agent would interpret the doctor’s notes on tumor shrinkage and side effects like nausea or low blood cell counts, and optimize the treatment dosage or schedule to maximize efficacy while reducing negative impacts on the patient’s health.
We can define the following connectors:
- Observation = Patient information (tumor size, blood work results such as white blood cell count, fatigue levels, nausea, overall well-being.). The patient information is obtained using a LLM tool that takes as input the doctor’s notes of the medical appointment. Additional sources could also be considered here (such as a databases of blood/scan results).
- Reward = reduction of tumor size and patient side effects. Those data are part of the observation and are obtained with the same LLM connector (see above).
- Action = chemotherapy dosage and schedule until the next appointment. A basic email connector for giving the action to the medical staff could be used here. If these emails are very basic (“Action recommended by the Adaptive cancer treatment agent on {day/month/date} for patient {name}: {action}”), a LLM isn’t even necessary here.
We have seen in this use case that a LLM tool would be used in two of the connectors: the observation and reward connectors.