Death past times suicide is a preventable tragedy if the suicidal private is identified too receives appropriate treatment. Unfortunately, some suicidal individuals produce non signal their intent, too others produce non have essential assistance. Youths alongside severe suicidal ideation are non taken seriously inwards many cases, too so are non admitted to emergency rooms. Influenza A virus subtype H5N1 mutual scenario is that resources are scarce, the ER is backed up, too a cursory clinical assessment volition create upward one's hear who is admitted too who volition endure triaged. From a practical standpoint, using fMRI to create upward one's hear suicide opportunity is a non-starter.
Yet hither nosotros are, alongside media coverage blaring that an Algorithm tin hand notice seat suicidal people using encephalon scans too Brain Patterns May Predict People At Risk Of Suicide. These media pieces herald a novel report claiming that fMRI tin hand notice predict suicidal ideation alongside 91% accuracy (Just et al. 2017). The authors applied a complex algorithm (machine learning) to analyze encephalon scans obtained using a highly specialized protocol to examine semantic too emotional responses to life too decease concepts.
Let me unpack that a bit. The scans of 17 immature adults alongside suicidal ideation (thoughts almost suicide) were compared to those from some other 17 participants without suicidal ideation. Influenza A virus subtype H5N1 calculator algorithm (Gaussian Naive Bayes) was trained on the neural responses to death-related too suicide-related words, too correctly classified xv out of 17 suicidal ideators (88% sensitivity) too xvi out of 17 controls (94% specificity). Are these results likewise goodness to endure true? Yes, probably. And however they're not goodness enough, because 2 at-risk individuals were non picked up.
The computational methods used to form out the suicidal vs. command groups are suspect, according to many machine learning experts on social media. One occupation is known every bit “overfitting” — using likewise many parameters taken from pocket-size populations that may non generalize to unique samples. The fundamental metric is whether the algorithm volition endure able to form out individuals from independent, out-of-sample populations. And nosotros don't know that for sure. Another occupation is that the leave-one-out cross validation is problematic. I'm non an practiced here, too so the Twitter threads that start below (and here) are your best bet.
ML re suicide, 90% correct, 2 groups of 17. Shiny journal. Anyone come across whatever problems ? https://t.co/mgQ8tW6s5w @tyrell_turing— KordingLab (@KordingLab) October 31, 2017
For the residuum of this post, I'll heighten other issues almost this report that concerned me.
Why role an expensive technology scientific discipline inwards the offset place?
The rationale for this included some questionable statements.
- ...predictions past times both clinicians too patients of time to come suicide opportunity have got been shown to endure relatively misfortunate predictors of time to come suicide attempt2,3.
...the implicit association of death/suicide alongside self was associated alongside an some 6-fold growth inwards the odds of making a suicide endeavour inwards the adjacent 6 months, exceeding the predictive validity of known opportunity factors (e.g., depression, suicide-attempt history) too both patients’ too clinicians’ predictions.But let's become ahead alongside an fMRI report that volition endure far to a greater extent than accurate than a brusque too easy-to-administer computerized test!
- Nearly 80% of patients who buy the farm past times suicide deny suicidal ideation inwards their final contact alongside a mental healthcare professional4.
How produce you lot stair out the neural correlates of suicidal thoughts?
This is a tough one, simply the authors suggest to uncover the neural signatures of specific concepts, every bit good every bit the emotions they evoke:
...the neural signature of the attempt concepts was treated every bit a decomposable biomarker of persuasion processes that tin hand notice endure used to pinpoint particular components of the change [in participants alongside suicidal ideation]. This decomposition attempts to specify a particular factor of the neural signature that is altered, namely, the emotional component...
How produce you lot direct which concepts too emotions to measure?
The “concepts” were words from 3 dissimilar categories (although the designation of Suicide vs. Negative seems arbitrary for some of the stimuli). The fix of xxx words was presented 6 times, alongside each discussion shown for three seconds followed past times a 4 minute blank screen. Subjects were “asked to actively mean value almost the concepts ... spell they were displayed, thinking almost their primary properties (and filling inwards details that come upward to mind) too attempting consistency across presentations.”
The “emotion signatures” were derived from a prior report (Kassam et al., 2013) that asked method actors to self-induce nine emotional states (anger, disgust, envy, fear, happiness, lust, pride, sadness, too shame). The emotional states selected for the introduce report were anger, pride, sadness, too shame (all chosen post hoc). Should nosotros hold off emotion signatures that are self-induced past times actors to endure the same every bit emotion signatures that are evoked past times words? Should nosotros hold off a universal emotional reply to Comfort or Evil or Apathy?
Six words (death, carefree, good, cruelty, praise, too trouble — inwards descending order) too v encephalon regions (left superior medial frontal, medial frontal/anterior cingulate, right middle temporal, left inferior parietal, too left inferior frontal) from a whole-brain analysis (that excluded bilateral occipital lobes for some reason) provided the most accurate discrimination betwixt the 2 groups. Why these specific words too voxels? Twenty-five voxels, specifically. It doesn't matter.
The neural representation of each concept, every bit used past times the classifier, consisted of the hateful activation flat of the v most stable voxels inwards each of the v most discriminating locations....and...
All of these regions, specially the left superior medial frontal surface area too medial frontal/anterior cingulate, have got repeatedly been strongly associated alongside self-referential thought......and...
...the concept of ‘death’ evoked to a greater extent than shame, whereas the concept of ‘trouble’ evoked to a greater extent than sadness inwards the suicidal ideator group. ‘Trouble’ also evoked less anger inwards the suicidal ideator grouping than inwards the command group. The positive concept ‘carefree’ evoked less pride inwards the suicidal ideator group. This pattern of differences inwards emotional reply suggests that the altered perspective inwards suicidal ideation may reverberate a resigned credence of a electrical current or time to come negative state of affairs, manifested past times listlessness, defeat too a grade of anhedonia (less pride evoked inwards the concept of ‘carefree’) [why non less pride to 'praise' or 'superior'? who knows...]
Not that this involves circularity or reverse inference or HARKing or anything...
How tin hand notice a method that excludes information from 55% of the target participants endure useful??
This i seems similar a showstopper. Influenza A virus subtype H5N1 amount of 38 suicidal participants were scanned, simply those who did non demo the desired semantic effects were excluded due to “poor information quality”:
The neurosemantic analyses ... are based on 34 participants, 17 participants per grouping whose fMRI information character was sufficient for accurate (normalized rank accuracy > 0.6) identification of the xxx private concepts from their fMRI signatures. The option of participants included inwards the primary analyses was based exclusively on the technical character of the fMRI data. The information character was assessed inwards price of the powerfulness of a classifier to seat which of the xxx private concepts they were thinking almost alongside a rank accuracy of at to the lowest degree 0.6, based on the neural signatures evoked past times the concepts. The participants who met this touchstone also showed less caput displace (t(77) = 2.73, P < 0.01). The touchstone was non based on grouping discriminability.
This logic seems round out to me, despite the claim that inclusion wasn't based on grouping classification accuracy. Seriously, if you lot throw out over one-half of your subjects, how tin hand notice your method always endure useful? Nonetheless, the 21 “poor information quality” ideators alongside excessive caput displace too bad semantic signatures were used inwards an out-of-sample analysis that also revealed relatively high classification accuracy (87%) compared to the information from the same 17 “good” controls (the information from 24 “bad” controls were excluded, apparently).
We attribute the suboptimal fMRI information character (inaccurate concept identification from its neural signature) of the excluded participants to some combination of excessive caput displace too an inability to sustain attending to the describe of piece of work of repeatedly thinking almost each stimulus concept for 3 s over a 30-min testing period.
Furthermore, some other classifier was fifty-fifty to a greater extent than accurate (94%) inwards discriminating betwixt suicidal ideators who had made a suicide endeavour (n=9) from those who had non (n=8), although the out-of-sample accuracy for the excluded 21 was exclusively 61%. Perhaps I'm misunderstanding something here, simply I'm puzzled...
I commend the authors for studying a neglected clinical group, simply want they were to a greater extent than rigorous, didn't overinterpret their results, too didn't overhype the miracle of machine learning.
Crisis Text Line [741741 inwards the US] uses machine learning to prioritize their telephone yell upward charge based on discussion usage too emojis. There is a swell variety of intersectional opportunity factors that may Pb someone to decease past times suicide. At present, no method tin hand notice capture the amount ambit of multifariousness of who volition cross the line.
If you lot are feeling suicidal or know someone who powerfulness be, hither is a link to a directory of online too mobile suicide assist services.
1 I won't hash out the problematic nature of the IAT here.
Just MA, Pan L, Cherkassky VL, McMakin DL, Cha c, Nock MK, & Brent D (2017). Machine learning of neural representations of suicide too emotion concepts identifies suicidal youth. Nature Human Behaviour. Published online: xxx Oct 2017
Kassam KS, Markey AR, Cherkassky VL, Loewenstein G, Just MA. (2013). Identifying Emotions on the Basis of Neural Activation. PLoS One. 8(6):e66032.
Nock MK, Park JM, Finn CT, Deliberto TL, Dour HJ, Banaji MR. (2010). Measuring the suicidal mind: implicit knowledge predicts suicidal behavior. Psychol Sci. 21(4):511-7. Sumber http://neurocritic.blogspot.com/
Can A Calculator Algorithm Pose Suicidal People From Encephalon Scans? The Answer Won't Surprise You
4/ 5Oleh Zeageat