In the development of prediction models, unsupervised Ai/ML systems are particularly susceptible to over fitting, tailoring to the build data set, and are insufficiently robust when tested on live, and not strictly parameter controlled, inputs. This is acutely seen in IVF Embryo selection tests, where group classification of success and failure can be inaccurate and are blurred. i.e. failure can be non-implantation, non-implantation due endometrial receptivity, biochemical pregnancy and first-trimester implantation loss. Based on our previous studies of metabolomic analysis and utilisation in Ai/ML prediction model algorithms, mass peak feature variables in spent blastocyst medium (SBM) predict which IVF blastocyst-embryo had the greatest chance of success. To operationalise, we took a probabilistic based prediction algorithm approach to create a simpler and more informative scoring system. Generated from a cohort of 385 individual MALDI-ToF Mass Spectrometry SBM sample spectra, we developed a more statically robust Bayesian based algorithm approach for a preimplantation metabolomic test for blastocyst metabolic competence (PMT-BC). This system results in direct input of analysed spectra, to output clinically informative probability scores and in a stratified probability format requested by embryologists. Visual user friendly, results are displayed in the form of a battery charge scaling and relates to metabolic systems being found in optimum ratio’s. The more metabolite ratio/systems found in the optimum the greater the probability of viable implantation. This opens up accessible metabolic profiling to the non-biochemist embryologist and implementation of metabolomic, PMT-BC, testing in the Clinical IVF centre. Furthermore, failure of high PMT-BC scoring IVF Blastocyst-Embryos will help define and quantify, failed fertility due to endometrial receptivity issues.
Iles et al. (Thu,) studied this question.