Adaptive feature selection and risk prediction for mental health decision support

  • Nawal Zaher

    Student thesis: Doctoral ThesisDoctor of Philosophy


    Mathematical modelling of mental health risks is a laborious task, since assessmentrecords contain diverse combinations of variables and a huge amount of missingdata. In addition, risk judgements made by assessors are not clearly formulatedfrom the available variables. The problem consists of two parts: first, selecting themost appropriate variables and, second, predicting risk using these variables, in realtime and in a manner that is clinically explainable.In this thesis, an adaptive feature selection algorithm is proposed based on MinimumRedundancy Maximum Relevance (MRMR) that builds on and extends theDynamic Feature Selection and Prediction (DFSP) algorithm [1]. A feed forwardapproach is utilised to reduce computational complexity. The selected features areused to predict risk through a linear regression model. The predictions are linearlyadjusted and unequal variance decision boundaries are used to handle heteroscedasticity.Two preprocessing steps are applied to reduce dimensionality and redundancy.The proposed algorithm is called Adaptive Feature Selection and Prediction(AFSP) and a method to autonomously update all its parameters, is devised.When the algorithm is applied to suicide risk prediction, the results show that AFSPhas better prediction accuracy than its predecessor, DFSP. The results also highlightthe improvement in accuracy and/or speed introduced by each component ofAFSP.The algorithm is also applied to two sub-concepts within suicide risk. First, AFSPis used to determine whether the absence of current intention stated by patients isreliable or not. Second, AFSP is used to predict patients that are in an episode ofclinical depression. The results of current intention and depression prediction arestatistically significant.The algorithm is intended to provide mental-health practitioners with prediction advicein real time, selecting the best factors for explaining predictions and updatingparameters autonomously off line so that they reflect the latest data.
    Date of Award27 Jun 2019
    Original languageEnglish
    SupervisorChristopher Buckingham (Supervisor) & George Vogiatzis (Supervisor)


    • risk assessment
    • suicide risk
    • suicidal intent
    • missing data

    Cite this