Predictive Models for Suicide Attempts in Major Depressive Disorder and the Contribution of EPHX2: A Pilot Integrative Machine Learning Study

Figure 5

The decision plots of ML models for four randomly selected research objects from the validation set. (a) An MDD case, due to relatively high CTQ subscale-emotional neglect, moderate ANT-executive control, high BIS subscale-cognitive impulsivity, and no other protective factors, D-Model B considered the probability of this patient being MDD to be 99.8%. (b) An HV case, due to relatively low BIS-total score, low CTQ subscale-emotional neglect, and low BIS subscale-cognitive impulsivity, although moderate ANT-executive control accounted for a small portion of the weight, D-Model B considered the probability of this patient being HV to be 77.2%. (c) A DSA case, due to relatively high HAMD-total score, high duration of current episode, and moderate 2-back-RT, although high age at onset accounted for a portion of the weight, S-Model C considered the probability of this patient being DSA to be 93.5%. (d) A DNS case, due to relatively low 2-back-RT, high ANT-alerting, and high ANT-mean ACC, although the role of several risk factors, S-Model C considered the probability of this patient being DNS to be 91.1%. This method could help users better understand the operation and decision-making process of ML models and thus timely intervene in the personalised risk factors of patients.

(a)

(b)

(c)

(d)