That PHQ-9 score doesn’t lie.

The HADS score confirms it.

And the paracetamol overdose last week removes any lingering doubt.

No problemo! A little CBT and an antidepressant should do the trick.

We have a whole candy store to choose from. Sertraline or fluoxetine. Citalopram… or oohh, maybe mitarzapine 😋. It’s all well and good having a shelf full of medications to pick from.
The problem, of course, is the picking.

You see, psychiatric management has a problem.

It’s all kinda based on vibes.
Why do we use lithium for bipolar disorder?

What exactly is going on in the mind of one with schizophrenia?

I don’t know. You don’t know. The psychiatrist is scratching their head too!

As a result, effective management is frustratingly inconsistent.

Only 20–25% of patients achieve remission with the first antidepressant prescribed. Roughly half experience minimal to no improvement at all. It’s no wonder many patients eventually think to hell with this and stop taking their medication altogether.

The trial-and-error system wasn’t working. Psych has been practically begging for a better method to this medical madness.

Enter PETRUSHKA: A (surprise surprise) AI-based prediction tool developed by the University of Oxford. This robot combines clinical evidence with patient preferences and a bunch of other data points. Slaps them all together and spits out a personalised antidepressant recommendation.

But this isn’t the first time a fancy tool with big promises has stepped foot in the wild west of psychiatry.

But before any shiny new tool gets taken seriously, it has to face the local sheriff: a randomised clinical trial.

So the aim of this RCT, published in , was to test whether using the PETRUSHKA tool to select an antidepressant, compared with clinician-led usual care:

Reduced rate of treatment discontinuation
Improves outcomes in adults with major depressive disorder

They recruited 520 eligible patients for the trial from 47 sites across Brazil, Canada and the UK. Mean starting scores ? PHQ-9 16.6, HAM-D = 16.3 and GAD-7 = 11.5.

The patients were then followed up at 4, 8 and 24 weeks, with the primary endpoint being all‑cause discontinuation of the initially prescribed antidepressant at 8 weeks (switching drug or stopping counted as discontinuation).

What did they find?

Well… it worked!
Patients whose antidepressants were selected using the PETRUSHKA tool were significantly less likely to stop their medication within the first 8 weeks(Control discontinuation 27% vs intervention discontinuation 17%). A roughly 38% relative risk reduction

Not only that:

Discontinuation due to adverse events at 8 weeks: 9% vs 16% [RR 0.59 (p=0.04)]

PHQ-9 score at 24 weeks: 7.1 vs 9.2 [adjusted mean difference of −1.92 (p<0.001)]

GAD-7 score at 24 weeks: 4.6 vs 5.8 [adjusted mean difference of −1.39 (p=0.002)]

Even down to the drug selection, differences were apparent. Whilst the clinician favoured sertraline (52%), PETRUSHKA recommended a wider variety of meds. Mirtazapine (29%), escitalopram (28%), and vortioxetine (24%) were the top choices.

Impressively, this is the first time a mental health clinical prediction tool has been demonstrated to be effective, according to the researchers.

Further research is needed to explore long-term outcomes and cost-effectiveness, but in this battle of doctor vs AI, AI takes the advantage.