Machine learning in the prediction of sperm motility

Published 21 November 2019

A new study from Norway suggests that machine learning via multimodal image understanding may improve on manual methods of semen analysis. Michael Riegler, a young ambassador for ESHRE and member of the Oslo research group, proposes a wider role for machine learning in male infertility.

Automatic analysis of different types of clinical data is currently advancing rapidly, in particular, multimodal image analysis (learning simultaneously from various sources of data). At this year's ESHRE Annual Meeting, for example, there were several presentations on the subject of machine learning (a subfield of artificial intelligence) and reproductive outcomes. Though promising, most of such current research in human reproduction is, from a machine learning point of view, still in its infancy.

Now, a new study from our group in Oslo shows that advanced machine learning methods for analysing videos of semen samples may be a useful tool in the investigation of male infertility.(1)

Manual semen analysis is central to male infertility investigation, but is time-consuming and requires extensive training to obtain reproducible results. Automatic analysis began in the 1980s but was very challenging because of factors such as background noise from other types of cells or particles in the video recordings. But now our study has demonstrated that a more sophisticated machine learning method is promising in automatic semen analysis for predicting sperm motility.

Our multimodal study analysed microscopic videos of 85 participants' semen samples and patient-related data. The latter was limited to readily available information including age, body mass index (BMI) and days of sexual abstinence. The aim was to determine whether the inclusion of these personal data could aid the prediction of the percentage of progressive and non-progressive sperm motility and immotile spermatozoa.

The results indicated that our selected deep learning algorithms did not lose or gain predictive power, even when sperm concentration was included in the analysis, in contrast to the computer-aided sperm analysis (CASA) systems where concentration is known to be a cofounding variable. Furthermore, we found that incorporating the time from collection to analysis, which inevitably influences sperm motility, represents an important advantage over all classical machine learning methods. The best method outperformed the baseline (average motility of the dataset as prediction, also called null model or ZeroR baseline) by an average mean absolute error of 4.20% for the prediction of motility. Importantly, our method was able to perform the prediction in five minutes including sample preparation, in contrast to the lengthy manual analysis.

Our work emphasises that evaluation of machine learning algorithms should be performed with care, as high evaluation metric scores are often not an indicator of a robust and well working algorithm. In addition, machine learning methods should not be evaluated on the same data used to train the models. Cross-validation is a data-efficient approach to avoid this and should be included in every analysis of these types of methods.

Although our study is the first to follow a multimodal approach, inclusion of various participant parameters in addition to video recordings did not improve - as envisaged - the prediction of sperm motility. This may be due to the type of data that was included (BMI and age); indeed, we have previously showed with other participant-related data that a multimodal analysis can lead to interesting results.(2,3). Machine learning may improve semen analysis, and notably as a less resource-demanding method than manual analysis. Future research should explore whether additional participant data - such as fatty acids, genomics or activity level - could be used and how it should be combined in multimodal analysis to increase the predictive power.

Finally, it is worth mentioning that our study was conducted on a dataset that is publicly available, allowing others to reproduce these results and perform other analyses on this topic.(4)

Overall, the findings indicate that machine learning opens a wide range of possibilities within the field of human reproduction. Furthermore, the quality and thoroughness of the evaluation of these methods should be considered at a high standard by the whole community to avoid a myriad of results of limited relevance, as can now be seen in other fields caught up in the midst of an 'artificial intelligence hype'.(5)

1. Hicks SA, Andersen JM, Witczak O, et al. Machine learning-based analysis of sperm videos and participant data for male fertility prediction. Nature Sci Rep 2019; 9, 16770. doi:10.1038/s41598-019-53217-y2.

2. Riegler MA, Andersen JM, Hammer HL, et al. Artificial intelligence as a tool in predicting sperm motility and morphology. Hum Reprod 2019; 34: suppl 1, P-116.

3. Witczak O, Andersen JM, Hicks SA, et al. Artificial intelligence predicts sperm motility from sperm fatty acids. Hum Reprod 2019; 34: suppl 1, P-120.

4. Haugen TB, Hicks SA, Andersen JM, et al. VISEM: a multimodal video dataset of human spermatozoa. In Proceedings of the 10th ACM Multimedia Systems Conference 2019: pp 261-266).

5. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 2019; 25: 44-56.