(PS5-C61) Attitudes Toward Machine Learning for the Identification of Suicide Risk

Attitudes Toward Machine Learning for the Identification of Suicide Risk

Frederick T. Schubert, Anna R. Gai, Kelsey L. Lowman, Thomas E. Joiner Jr.

Department of Psychology, Florida State University, Tallahassee FL

Background: Machine learning decision support systems offer a revolutionary approach for healthcare systems to identify those at high risk for suicide and other negative medical outcomes. While this approach shows promise in providing easier outlets for individuals at high risk to receive help, identification by technology may actually be perceived as aversive and impede help-seeking. Additionally, prior research suggests that preexisting factors such as healthcare system distrust and political affiliation may influence mental health stigma and help-seeking. Thus, it is important to understand attitudes toward the use of machine learning algorithms for identifying suicide risk (ML-SI). This study sought to evaluate attitudes toward ML-SI among the general public as well as individual factors related to these attitudes.

Method: Participants (N = 250) were recruited through Amazon’s Mechanical Turk (59.4% male, M/SD age = 32.93/9.06 yrs). Five items assessing participants’ attitudes toward ML-SI were rated from 0-100 and summed for an overall ML-SI Attitudes score, with higher scores indicating more positive attitudes. Participants also completed measures assessing political affiliation (PA), self-stigma of seeking psychological help (SS), and health-care system distrust (HCD), where higher scores indicated more conservatism, stigma, and distrust, respectively.

Results: On average, ML-SI Attitudes were rated 61.84 out of 100 (SD = 18.95). At the bivariate level, ML-SI Attitudes were significantly positively associated with PA, SS, and HCD (rs = .41, .40, .25, respectively, ps < .001). However, when ML-SI Attitudes were regressed onto all three variables simultaneously, only PA and SS remained significant (βs = .31, ps < .001), such that those with more self-stigma and more conservative political affiliations had more positive ML-SI Attitudes. Given the association between PA and SS observed in the current study (r = .35, p < .001) and in the literature, we tested for an interaction between the two in predicting ML-SI Attitudes. The interaction was significant (β = .14, p = .01), such that the positive relationship between SS and ML-SI Attitudes was stronger at higher levels of PA (i.e. conservatism; β = .42, p < .001), compared to lower levels of PA (i.e. liberalism; β = .16, p = .047).

Conclusions: Results indicate that general attitudes toward ML-SI may be closer to neutral and not overtly positive. Additionally, the associations between ML-SI, PA, and SS suggest that self-stigma is more strongly associated with ML-SI attitudes for those who are politically conservative. These findings carry important implications for identifying individuals for who may respond positively to ML-SI identification as well as those who may respond negatively. The exploratory nature of these analyses may limit the generalizability and power of these findings. Future research should elucidate how more granular measures of political affiliation are related to ML-SI and identify whether the acceptability of ML-SI can be improved.