Using NLP to detect mental health crises

I work in patient safety and suicide prevention and people always mangle the statistics.

Between 2000–2020, rates of suicide increased worldwide1

This coincides with the WHO effort to reduce deaths by suicide, and that programme includes a lot of work to improve the data. Are more people dying by suicide, or are more countries counting deaths better? The article cites this WHO report: Suicide World Wide in 2019

That document says this:

As with previous revisions of WHO Global Health
Estimates, the entire time series from the year
2000 were revised, including for suicide. Because
the estimates for the years 2000–2019 draw on
new data and on the results of the GBD2019
study, and because there have been substantial
revisions to the methods used, these estimates
are not comparable with previous estimates for
2000–2016 or with earlier revisions published by
WHO. Therefore, suicide estimates in this booklet
supersede suicide estimates previously published
by WHO, and differences between published
revisions should not be interpreted as time trends.6
Rather, to identify time trends in WHO Global
Health Estimates, including suicide, one should
look at the latest published revision of WHO Global
Health Estimates (Figure 10).

And in figure 10 it says this:

In the 20 years between 2000 and 2019, the global age-standardized suicide rate decreased by 36%, with
decreases ranging from 17% in the Eastern Mediterranean Region to 47% in the European Region and
49% in the Western Pacific Region (Figure 10). The only increase in age-standardized suicide rates was
in the Region of the Americas, reaching 17% in the same time-period. The global rate also decreased for
age-group specific rates.

With regard to the SDGs, a global acceleration of the decrease in the suicide mortality rate is needed to
reach the global target of a one-third reduction by 2030 that countries have committed to

I'm not sure how a WHO document saying "the global age-standardized suicide rate decreased by 36%" is compatible with the article's claim of "Between 2000–2020, rates of suicide increased worldwide".

The tech looks interesting, but I'll note that previous attempts to create risk ratings for suicidality have been so bad the UK NICE (looks at evidence bases for medical interventions) lists it as a "Do not use". (I understand these are very different types of tool.)

NICE - Self-Harm: assessment, management and preventing recurrence - 1.6 Risk assessment tools and scales

This is super cool. It's a relief to see NLP being used in this domain in a way that is helps competent human operators rather than replaces them, and from my quick look at the journal article it seems like they put the amount of thought into how they defined and labelled their classes that it deserves given such a domain.