Background: The current rapid expansion of commercially-available tests to assess the microbiome and metabolome provides an exciting backdrop for the development of personalised and preventative approaches to health. However, the application of these tests is both geographically and economically restrictive; they are generally limited to the country in which the laboratory resides, and likely to remain prohibitively expensive for the majority of the population for a number of years. Machine learning provides an ideal platform to predict complex test results from more subjective (and widely-available) data, particularly when restricted to well-defined patient groups.

Methods: XGBoost was trained with data from 802 athletes (501 males, 301 females) with common health complaints. This included gastrointestinal (GI) symptoms, depression or anxiety, weight gain, poor recovery, insomnia, and fatigue. All participants completed a standardised health assessment questionnaire (HAQ), answering 53 quality-of-life questions rated on an analog scale. Biochemical data was assessed using commercially-available stool microbiome and culture tests, urinary metabolome tests (organic acids, cortisol and sex hormone metabolites), and a blood biochemistry panel. The Synthetic Minority Over-sampling Technique (SMOTE) was used to balance classes in the training dataset. The algorithm was then tested on 20% of the participant data (initially withheld) to assess its ability to predict pre-determined biochemical patterns based purely on responses to the HAQ. The patterns assessed were: 1) blood glucose dysregulation (elevated fasting blood glucose, fasting insulin, or HbA1c), 2) low haemoglobin 3) low sex hormones (testosterone in males, oestrogen in females), 4) the presence of opportunistic GI pathogens, and 5) dysregulation of diurnal cortisol rhythm (from a four-point cortisol assessment).


Results: Based on the testing data, blood glucose dysregulation (80% sensitivity, 98% specificity), low haemoglobin (94-100% sensitivity, 97-98% specificity), low sex hormone levels (83-97% sensitivity, 100% specificity), presence of GI pathogens (82-98% sensitivity, 98-100% specificity), and dysregulation of diurnal cortisol rhythm (87% sensitivity, 100% specificity) could all be predicted from HAQ responses with a high degree of accuracy. Ranges of sensitivity and specificity occurred as a result of including multiple predictions in a single category, or if the target reference ranges were different for males and females.

Discussion: In this well-defined population of athletes, the XGBoost algorithm was able to predict the presence of common biochemical patterns from subjective questionnaire data with levels of sensitivity and specificity approaching what would be expected from a “gold standard” medical test. Combining machine learning with both subjective and complex metabolomic and microbiome data can aid the clinician in applying more targeted testing and treatment, reducing the economic burden on the patient and healthcare system, and increasing the number of people that can gain access to personalised and preventative health care approaches.


Author: Thomas Wood

Coauthor(s): Christopher Kelly BSc

Status: Completed Work