Nils Hentati Isacsson

PhD student at Karolinska institutet. Working with machine learning in internet psychiatry.

[Nils Hentati Isacsson @H_I_now]

7/6 Oh and a synthetic dataset is available in the supplement for analysis-reproducibility and exploration with identical properties as the original dataset. Thanks #RStats #synthpopR

[Nils Hentati Isacsson @H_I_now]

5/6 Turns out that if you look at each symptom separately it is the rule rather than the exception that patients have some symptoms that increase, some that decrease, and some that stays the same throughout treatment - despite the sum decreasing.

[Nils Hentati Isacsson @H_I_now]

6/6 Albeit a very simple illustration of the caveats of only looking at the sum of symptom severities - it underscores the problem with the current definitions of treatment outcomes and current measurement practices.

[Nils Hentati Isacsson @H_I_now]

4/6. However, since patients usually have a heterogenous symptom picture (and measurements usually have varying validity/reliability)- how do each of these symptom change even though we "think" the patient to have improved by looking at the sum/category?

[Nils Hentati Isacsson @H_I_now]

3/6 The usual way to measure symptom reduction is to sum all the symptoms into a representative score - or committing a sin with a binary category of this score. And then evaluate treatments based on the means of these sums/ these categories.

[Nils Hentati Isacsson @H_I_now]

2/6 New Preprint about the possible mean Means when looking at- and defining- treatment outcomes in psychological treatments. Summary: We know patients in psychiatric care vary. This makes measurement of symptoms within treatments a challenge.

[Nils Hentati Isacsson @H_I_now]

1/6 It is the rule rather than the exception that patients fluctuate in symptoms during treatment for anxiety and depression 👉 focusing only on the sum of symptoms misses this heterogeneity. https://t.co/LBshSwWX4I #measurement #psychotherapy #Psychiatry

[Nils Hentati Isacsson @H_I_now]

@lluaces @krstoffr haha yeah. don't skip the baselines. Along those lines I like: Christodoulou, et al., (2019). A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. https://t.co/wqHABeKRk2

[Karandeep Singh @kdpsinghlab]

The presenter puts up a slide showing “random forest variable importance.” You know the one...

The sideways bar plot.

Says “only showing the top 20 variables here...” to highlight the hi-dimensional power of random forests.

The slide is awkwardly wide-screen. Everyone squints.

[Nils Hentati Isacsson @H_I_now]

@dingding_peng Many thanks for this !!

[Nils Hentati Isacsson @H_I_now]

@dingding_peng May I ask why it counters ceiling effects? / any links for the ignorant?

[Gustav Nilsonne @GustavNilsonne]

A paper last week claimed to show that an EEG-based model could predict response to antidepressant treatment. Important, if true. So I took a closer look and reanalysed some of their data. The model doesn't work. Here's the story: https://t.co/RlO37EBkaY (1/x)

[Nils Hentati Isacsson @H_I_now]

@krstoffr Seen this ? Hanel, P. H. P., Maio, G. R., & Manstead, A. S. R. (2019). A new way to look at the data: Similarities between groups of people are large and important. Journal of Personality and Social Psychology, 116(4), 541–562. https://t.co/8QPXImBpjU

[Nils Hentati Isacsson @H_I_now]

#psyTeachR Trying to follow your great resources however: https://t.co/MdjKyut7au is empty - working as intended (I have to find my own open dataset) or broken link?)

[Nils Hentati Isacsson @H_I_now]

@MaartenvSmeden Must be common practice?

[Farzad Mostashari @Farzad_MD]

1/ #EpiTwitter tweetstorm coming

Warning: strong opinions

The analysis is naive and the findings are ridiculous; the fact that it was published is a sign that when medical journal editors hear "deep learning AI" their brains stop working.

[ESRII 2019 Copenhagen @ESRII2019CPH]

Should we let THE MACHINE solve everything? @IsacssonNils continues the discussion on algorithmic predictions in Internet-based interventions #ESRII2019

[Kareem Carr, Ph.D. @kareem_carr]

lawyer: and what percent of time did you spend fitting the model?
me, a data scientist: 2%
lawyer: and the other 98%?!
me: cleaning the data
lawyer: CLEANING THE DATA!!!
judge: *gasp*
jury: *gasp*
audience:*gasp*
me: Well ya know what guys, I’m not super happy about it either!

[Kristoffer Magnusson @krstoffr]

New post: Change over time is not "treatment response"

The post includes a simple simulation of why differences in change over time is not the causal estimand we want when looking for (non-)responders.

https://t.co/XEKTqrk9Lq