Response-to-Treatment as Diagnostic Test¶
Why It Matters¶
Traditional thinking separates diagnosis (figure out what's wrong) from treatment (fix it). But treatment responses provide information - if the patient improves with X, that makes conditions responsive to X more likely. This tool corrects the error of treating diagnosis and intervention as sequential rather than intertwined.
The Key Move¶
When facing diagnostic uncertainty, identify treatments that would work for some possibilities but not others. Administer the treatment and observe the response. Improvement shifts probability toward responsive conditions; lack of improvement shifts probability toward non-responsive conditions or incorrect treatment. The key: you must predefine what improvement looks like and how quickly you expect it.
Classic Application¶
Bronchodilator response in shortness of breath. A patient is short of breath - could be asthma exacerbation (responds to bronchodilators), heart failure (doesn't respond to bronchodilators), or pneumonia (doesn't respond). Give inhaled albuterol and reassess in 15 minutes. If breathing improves, asthma becomes much more likely; if unchanged, shift probability to cardiac or infectious causes. The treatment doubles as a diagnostic test.
Surprising Application¶
Engineering debugging by targeted intervention. A system is slow - could be database queries, network latency, or memory pressure. Add database query caching and observe. If latency drops, database was the bottleneck; if unchanged, it's elsewhere. The "treatment" (caching) provides diagnostic information about the underlying "condition" (bottleneck location). This is more efficient than passive observation when interventions are cheap and reversible.
Failure Modes¶
Confusing correlation with causation - improvement might be spontaneous rather than treatment-caused. This requires understanding natural history and timing of response. The tool fails when treatments have delayed effects (you need immediate feedback for diagnostic value) or when treatments have side effects that mask the diagnostic signal. Also fails catastrophically when the diagnostic treatment is itself harmful - giving blood thinners to "test if there's a clot" is dangerous if there's actually bleeding. Finally, placebo response can create false diagnostic information.
Go Deeper¶
Sox, H.C. et al. (2013). Medical Decision Making, 2nd ed., particularly Chapter 8 on the value of therapeutic trials; Rasouli, M. & Kalantari, K. (2010). "Comparison of Methods to Handle Missing Data on Mental Health Questionnaires." Epidemiology and Psychiatric Sciences, 19(3), 318-326.