Table 3 Research using ML to detect eating disorder risk via physiological predictors
Study | Sample | Predictors variables | Outcome variables | Best performing ML approach | Explanatory power of best performing model | Limitations |
|---|---|---|---|---|---|---|
Cerasa et al. [50] | 36 females, with an ED diagnosis (n = 17) or body mass index-matched healthy controls (n = 17) | Structural magnetic resonance images | ED status | Support vector machine | 85% | Small sample size |
Cyr et al. [52] | 84 female adolescents, who met criteria for BN (n = 28), subclinical BN (SBN; n = 16), or healthy controls (HC; n = 40) | Functional magnetic resonance images of fronto-striatal regions during performance of a Simon task | ED status | Support vector machine | BN v. HC = 58% SBN v. HC = 64% Train BN v. HC, Test SBN v. HC = 66% | Small sample |
Guo et al. [57] | 13,206 adolescent and adult females, who have AN (n = 3940) or healthy controls (n = 9266) | Whole genome genotyping data | AN status | Logistic regression with LASSO penalty | 69% |  |
Ioannidis et al. [70] | 3937 observations from 36 AN inpatients | Physiological parameters, blood investigations over a 1-year period | Medical risk defined by independent clinical rates of deteriorating cases | Random forest | 98% | Small sample |
Lavagnino et al. [55] | 30 females, with an AN diagnosis (n = 15) or demographically matched healthy controls (n = 15) | Structural neuroimaging scans | ED status | Least absolute shrinkage and selection operator (LASSO) | 83% | Small sample |
Lavagnino et al. [54] | 67 adult females, who have restrictive-type AN (n = 19), who have recovered from restrictive-type AN (n = 24), or healthy controls (n = 24) | Structural brain scans to test cortical thickness | ED status | Linear relevance vector machine | AN v. HC = 74% | Small sample |
Strigo et al. [53] | 1 adolescent female with mixed ED, depressive, and gastrointestinal symptoms | Functional magnetic resonance images during a pain anticipation program and psychological survey responses created on previous samples | Which diagnostic phenotype most closely approximates the patient | Support vector machine | 56% based on brain activation 84% based on psychological variables | Case study |
Weygandt et al. [51] | 70 females, who met criteria for binge eating disorder (BED; n = 17), or bulimia nervosa (BN; n = 14), or normal-weight controls (C-NW; n = 19) or overweight controls (C-OW; n = 17) | Functional imaging from a whole-body tomograph while viewing food or neutral images | ED status | Support vector machine | BED v. C-NW = 86% BN v. C-NW = 78% BED v. C-OW = 71% BN v. C-OW = 86% BED v. BN = 84% | Small sample size |