The split
Causal medical work divides naturally into two question types. Both involve patients, both involve interventions, both fail the same way when treated as prediction problems — but the questions they answer are distinct, and the methods they reach for differ accordingly.
Personalized Medicine assumes a treatment is on the table and asks how it will go for this particular patient. The trial averages do not transfer cleanly. The covariates that select patients into treatment are exactly the covariates that drive response. The relevant question is Rung 3: what would have happened, for this individual, under the alternative?
Clinical Decision-Making assumes a guideline exists and asks whether following it is doing harm. Many decisions are downstream of a protocol calibrated on a population average; many of those protocols, examined causally, turn out to harm the specific patient in front of the clinician. The relevant question is whether the intervention is justified at all, given what the model says will happen.
The two branches share their formal machinery — structural causal models, the do-operator, identification, counterfactual abduction — but apply it to different question shapes.
Personalized Medicine
Standard clinical statistics asks what works on average. Personalized Medicine asks what works for this patient — given their history, biomarkers, treatment trajectory, and the specific drugs that have already failed.
Five worked case studies in this branch:
- Pharmacovigilance — Adverse Event Attribution: did the drug cause the injury in this patient?
- Oncology — Immunotherapy vs. Chemotherapy: who benefits from which?
- Drug Repurposing — Transporting Trial Results to Novel Target Populations.
- Psychiatry — Treatment-Resistant Depression: counterfactual sequencing after failed trials.
- Critical Care — Sepsis: dynamic treatment regimes via causal reinforcement learning.
Clinical Decision-Making
Guidelines compress trial averages into protocols applied to every patient. When the protocol is calibrated for the wrong patient, the average correct decision becomes the individual wrong one. Clinical Decision-Making asks whether the treatment in front of you is doing harm.
Two worked case studies in this branch:
- Statins & Hospitalisation — The correct average applied to the wrong individual.
- Iatrogenic Medications — When the drug isn’t fighting the disease, it’s fighting the other medications.
Why two branches, not one
It would be possible to flatten medicine work into a single cluster — one list of seven case studies, all under "Medicine." The reason not to is that the questions answered on either side are differently shaped, and grouping them as if they were interchangeable obscures the framing a reader needs.
A patient’s oncologist is asking a Personalized Medicine question: given everything we know about this tumor, which treatment maximizes survival for this patient? A health system’s prescribing audit is asking a Clinical Decision-Making question: given the cascade of drugs this patient is now on, is the next one going to do net good? The first question requires individual-level abduction. The second requires intervention analysis on a system already in motion. Same toolkit, different operations.
Splitting them lets each branch develop its own internal coherence. A reader looking for "how to think about treatment for this patient" finds the five PM cases. A reader looking for "is this protocol harming this patient" finds the two CDM cases. Neither has to wade through the other’s material to find the relevant one.