Medical intelligence vs. artificial intelligence — How psychiatrists arrive at treatment algorithms

Abstract

The rise of artificial intelligence and computational methods in clinical decision-making allows for a novel perspective on personalized psychopharmacology. While treatment guidelines bring forth treatment algorithms, their recommendations do not take into account the complexities of clinical practice and patient individuality. As guidelines provide specifications at broader levels of prescription, they cannot contribute fully personalized predictions of treatment outcomes. Yet, professionals need to estimate treatment response when prescribing specific agents to individual patients. It is currently not well understood how psychiatrists bridge the gap between broad guideline recommendations and individual patient needs. We present a qualitative analysis from 16 senior German psychiatrists from seven clinical institutions. The analysis reveals the broad heuristics used by experienced clinicians to personalize treatment recommendations on the individual patient level. The analysis indicates that professionals follow a flexible hierarchy leading stepwise from diagnosis, symptoms, history, adverse medication effects, to comorbidity. At each step, clinicians consider individual factors such as age, gender, and patients’ preferences. With increasing complexity, guideline adherence weakens and the influence of clinical experience and intuition increases. Prescribers’ choices primarily derive from specialized training, their teachers, and clinical routine. The constraints to algorithm improvement are a lack of objective diagnostic procedures, of intercollegiate communication, and of courage to diverge from established practices. The results illuminate our sample’s treatment algorithm and its sources. We relate our findings to current discussion on treatment optimization through the involvement of AI in light of the potential harms of uncritical adoption of AI tools into psychiatric practice.