Table 1 Rules-of-thumb for most suitable machine learning algorithms

AI method	Learning type	Common tasks	Must suitable data types	Quantity of data required	Interpretability	Example use in ART?
Linear/Logistic regression	Supervised	C&R	Numerical	++	+++	Optimizing trigger day timing²⁷
Decision tree	Supervised	C&R	Numerical, categorical	++	+++	Decision-making during OS³⁰
k-NN	Supervised	C&R	Numerical, categorical	+	++	Optimizing starting dose during OS¹¹
SVM	Supervised	C&R	Numerical, categorical	++	++	Streamlining monitoring of patients during OS³¹
Random forest	Supervised	C&R	Numerical, categorical	++	++	Predicting risk of OHSS during OS³²
CNN	Supervised, unsupervised	C&R, clustering	Image, audio, text	+++	+	Predicting ploidy status of an embryo¹⁰⁰
k-means	Unsupervised	Clustering	Numerical	++	++	Effect of sperm parameters on IVF outcomes⁶⁴
GAN	Unsupervised	Generative	Image, time-series, text	+++	+	Generating synthetic embryo images⁷³
LLM	Unsupervised	Generative	Text	+++	+	Pre-treatment counseling⁶

Rules-of-thumb in determining the most suitable machine learning algorithm for a task with relevant examples of their application. Three plus signs imply the highest requirement or capacity, and one plus sign the lowest. For example, the convolutional neural network (CNN) supports several data types, and generally requires high quantities of data (i.e., thousands) for adequate performance, but exhibits poor interpretability (i.e., ‘black-box’). Conversely, k-nearest neighbors (k-NN) can work well even with only hundreds of data samples, and the weighting of predictors can be reasonably estimated for interpretability purposes. AI artificial intelligence, ART assisted reproductive technology, C&R classification and regression, SVM support vector machine, CNN convolutional neural network, GAN generative adversarial network, LLM large language model, OS ovarian stimulation, OHSS ovarian hyperstimulation syndrome.

Quick links

Search