Portal de Periódicos da CAPES

Precision medicine and artificial intelligence: overview and relevance to reproductive medicine

2020; Elsevier BV; Volume: 114; Issue: 5 Linguagem: Inglês

10.1016/j.fertnstert.2020.09.156

1556-5653

Iman Hajirasouliha, Olivier Elemento,

AI in cancer detection

Traditionally, new treatments have been developed for the population at large. Recently, large-scale genomic sequencing analyses have revealed tremendous genetic diversity between individuals. In diseases driven by genetic events such as cancer, genomic sequencing can unravel all the mutations that drive individual tumors. The ability to capture the genetic makeup of individual patients has led to the concept of precision medicine, a modern, technology-driven form of personalized medicine. Precision medicine matches each individual to the best treatment in a way that is tailored to his or her genetic uniqueness. To further personalize medicine, precision medicine increasingly incorporates and integrates data beyond genomics, such as epigenomics and metabolomics, as well as imaging. Increasingly, the robust use and integration of these modalities in precision medicine require the use of artificial intelligence and machine learning. This modern view of precision medicine, adopted early in certain areas of medicine such as cancer, has started to impact the field of reproductive medicine. Here we review the concepts and history of precision medicine and artificial intelligence, highlight their growing impact on reproductive medicine, and outline some of the challenges and limitations that these new fields have encountered in medicine. Traditionally, new treatments have been developed for the population at large. Recently, large-scale genomic sequencing analyses have revealed tremendous genetic diversity between individuals. In diseases driven by genetic events such as cancer, genomic sequencing can unravel all the mutations that drive individual tumors. The ability to capture the genetic makeup of individual patients has led to the concept of precision medicine, a modern, technology-driven form of personalized medicine. Precision medicine matches each individual to the best treatment in a way that is tailored to his or her genetic uniqueness. To further personalize medicine, precision medicine increasingly incorporates and integrates data beyond genomics, such as epigenomics and metabolomics, as well as imaging. Increasingly, the robust use and integration of these modalities in precision medicine require the use of artificial intelligence and machine learning. This modern view of precision medicine, adopted early in certain areas of medicine such as cancer, has started to impact the field of reproductive medicine. Here we review the concepts and history of precision medicine and artificial intelligence, highlight their growing impact on reproductive medicine, and outline some of the challenges and limitations that these new fields have encountered in medicine. Discuss: You can discuss this article with its authors and other readers at https://www.fertstertdialog.com/posts/31477 Discuss: You can discuss this article with its authors and other readers at https://www.fertstertdialog.com/posts/31477 Traditionally, new treatments have been developed for the population at large. In the past few decades, technological advances, such as the ability to genotype or sequence entire genomes, have revealed tremendous genetic diversity between individuals (1Auton A. Brooks L.D. Durbin R.M. Garrison E.P. Kang H.M. et al.1000 Genomes Project ConsortiumA global reference for human genetic variation.Nature. 2015; 526: 68-74Crossref PubMed Scopus (7682) Google Scholar). The same technologies can reveal mutations that patients with inherited disorders have acquired from their parents. In diseases driven by genetic events such as cancer, genomic sequencing can unravel all the mutations that drive individual tumors (2Rennert H. Eng K. Zhang T. Tan A. Xiang J. Romanel A. et al.Development and validation of a whole-exome sequencing test for simultaneous detection of point mutations, indels and copy-number alterations for precision cancer care.NPJ Genom Med. 2016; 1PubMed Google Scholar, 3Wrzeszczynski K.O. Felice V. Abhyankar A. Kozon L. Geiger H. Manaa D. et al.Analytical validation of clinical whole-genome and transcriptome sequencing of patient-derived tumors for reporting targetable variants in cancer.J Mol Diagn. 2018; 20: 822-835Abstract Full Text Full Text PDF PubMed Scopus (13) Google Scholar). This enhanced ability to unravel the genetic makeup of individual patients has led to the concept of precision medicine. Precision medicine unravels the genetic uniqueness of individuals and the singular molecular makeup of their disease and then seeks to use that information to match each individual to the best treatment. Such matching is increasingly possible because we now have access to a variety of treatments for many diseases and an increased understanding of the molecular alterations associated with the likely response to individual treatments. For example, cancer patients with mutations affecting the NTRK kinase respond particularly well to NTRK inhibitors (4Drilon A. Laetsch T.W. Kummar S. DuBois S.G. Lassen U.N. Demetri G.D. et al.Efficacy of Larotrectinib in TRK fusion-positive cancers in adults and children.N Engl J Med. 2018; 378: 731-739Crossref PubMed Scopus (1334) Google Scholar). These targeted agents are not only highly effective but also generally show less toxicity in other cells and tissues since those do not harbor the target mutations. The ability to query circulating tumor DNA instead of biopsy material may mean that genomic analysis will be increasingly available to more patients in the future (5Bettegowda C. Sausen M. Leary R.J. Kinde I. Wang Y. Agrawal N. et al.Detection of circulating tumor DNA in early- and late-stage human malignancies.Sci Transl Med. 2014; 6224ra24Crossref PubMed Scopus (2774) Google Scholar). Precision medicine research seeks to increase our ability to understand and predict the efficacy of treatment, given the information that can be obtained about the uniqueness of each individual. Increasingly, precision medicine extends beyond genomics and applies to epigenetics, proteomics, and metabolomics as well. In addition, quantifying environmental exposure (6Vermeulen R. Schymanski E.L. Barabasi A.L. Miller G.W. The exposome and health: where chemistry meets biology.Science. 2020; 367: 392-396Crossref PubMed Scopus (224) Google Scholar), behaviors (7Althoff T. Sosic R. Hicks J.L. King A.C. Delp S.L. Leskovec J. Large-scale physical activity data reveal worldwide activity inequality.Nature. 2017; 547: 336-339Crossref PubMed Scopus (447) Google Scholar), and the immune system (8Adlung L. Amit I. From the Human Cell Atlas to dynamic immune maps in human disease.Nat Rev Immunol. 2018; 18: 597-598Crossref PubMed Scopus (16) Google Scholar) can enhance precision medicine. Even in genomics, the ability to focus on individual cells (single-cell genomics) has unraveled heterogeneity not previously seen (9Izar B. Tirosh I. Stover E.H. Wakiro I. Cuoco M.S. Alter I. et al.A single-cell landscape of high-grade serous ovarian cancer.Nat Med. 2020; 26: 1271-1279Crossref Scopus (104) Google Scholar, 10Wagner J. Rapsomaniki M.A. Chevrier S. Anzeneder T. Langwieder C. Dykgers A. et al.A single-cell atlas of the tumor and immune ecosystem of human breast cancer.Cell. 2019; 177: 1330-1345.e18Abstract Full Text Full Text PDF PubMed Scopus (300) Google Scholar). These data confirm observations seen in patients, where targeted therapies can be effective for a short amount of time but patients develop resistance because cells may have preexisting resistance mutations in a large cell population (11Wagle N. Emery C. Berger M.F. Davis M.J. Sawyer A. Pochanard P. et al.Dissecting therapeutic resistance to RAF inhibition in melanoma by tumor genomic profiling.J Clin Oncol. 2011; 29: 3085-3096Crossref PubMed Scopus (764) Google Scholar). More and more, precision medicine is being enhanced by artificial intelligence (AI). Machine learning is a form of AI in which a machine can learn and adapt to situations and data training. Typically, a training data set is used to train a computer program to connect objects, such as images that are described using a series of features—color, shape, and texture—with specific labels or classes, for example, cancer or noncancer. Once trained, the computer program, called a classifier, is used to label objects. Machine learning encompasses two main approaches: supervised and unsupervised learning. In supervised learning, the classes or labels are known during training. In contrast, unsupervised learning, such as hierarchical clustering, is used to discover structure within data, such as the existence of classes. There are several machine learning techniques that can be used to learn how to map objects to classes and create predictive models. Some of the most frequently used techniques include logistic regression, random forests (RFs), naive Bayes classifier, support vector machines (SVMs), and artificial neural networks, including the recently developed deep neural networks. Logistic regression is a widely used supervised statistical technique for predictive analysis. It can explain the relationship between a binary variable (e.g., cancer or noncancer) and one or more independent variables (e.g., genetics, cigarette smoke). Logistic regression uses a sigmoid function to map the weighted sum of features describing an object to the probability that the object belongs to a particular class (12Berkson J. Application of the logistic function to bio-assay.J Am Stat Assoc. 1944; 39: 357-365Google Scholar). The weights are learned from data (many labeled objects) during training using the statistical technique of maximum likelihood estimation. Logistic regression is a popular machine learning technique for several reasons: [1] it is easy to implement and has few extra parameters that need to be tuned; [2] it provides intuitive interpretation of the features that contribute to prediction accuracy, including statistical assessment of their contribution in the form of P values; [3] it can be used to model interactions between variables; and [4] it generally demonstrates good accuracy when the mapping between classes and training data is not overly complex. Random forests connect objects to classes by building a large number of decision trees from randomly selected subsets of the training data. Each decision tree learns how to map objects to classes using simple rules. Each decision tree gets to “vote” its class on each object, and the sum of each vote is used to estimate the probability that an object belongs to a particular class. As such, RFs are often referred to as an ensemble method (13Amit Y. Geman D. Shape quantization and recognition with randomized trees.Neural Comput. 1997; 9: 1545-1588Crossref Scopus (765) Google Scholar). The use of decision trees allows RFs to capture complex interactions between variables, thus frequently enabling strong classification performances. Conveniently, RFs can also naturally handle both discrete and continuous features together in the same model, a situation frequently encountered in biomedical data sciences. Naive Bayes classifiers use the Bayes rule to estimate the probability that an object belongs to a particular class given data provided about the object (14Domingos P. Pazzani M. On the optimality of the simple Bayesian classifier under zero-one loss.Machine Learning. 1997; 29: 103-130Crossref Google Scholar). One key feature of naive Bayes classifiers (what makes them “naive”) is that each feature is considered to be independently predictive of the class. The weight of each feature is calculated from labeled data as part of the training. Naive Bayes classifiers are popular due to their intuitive, interpretable, and modular structure, which remains firmly grounded in the statistical field (15Madhukar N.S. Khade P.K. Huang L. Gayvert K. Galletti G. Stogniew M. et al.A Bayesian machine learning approach for drug target identification using diverse data types.Nat Commun. 2019; 10: 5221Crossref Scopus (71) Google Scholar). Naive Bayes classifiers use Bayes’s theorem in their decision rule and are thus very efficient and highly scalable. They are particularly useful for large data sets. Support vector machines learn to classify objects by identifying the hyperplane that best separates objects from classes being analyzed (16Boser BE, Guyon IM, Vapnik VN. A training algorithm for optimal margin classifiers. Proceedings of the 5th Annual ACM Workshop on Computational Learning Theory 1992:144–152.Google Scholar); by projecting input data into higher dimensional spaces, SVMs can learn to predict complex classes of objects. While SVMs are effective when dimensions outnumber samples, they can be slow and less applicable for large-scale data. Artificial neural networks are machine learning techniques that mimic aspects of how brains work (17Rosenblatt F. The perceptron: a probabilistic model for information storage and organization in the brain.Psychol Rev. 1958; 65: 386-408Crossref PubMed Scopus (4961) Google Scholar). A neural network typically consists of several layers of artificial neurons fully connected to each other by edges, each one associated with a weight. Each neuron receives signals from multiple neurons in the previous layer, integrates those signals, and “fires” if the integrated signals are above a specific threshold. Artificial neural networks learn using a technique of gradient descent, propagating and reducing classification errors from layers to layers. Artificial neural networks generally consists of three layers: input, hidden, and output. Deep neural networks extend neural networks by increasing the number of layers and the number of neurons per layer (18LeCun Y. Backpropagation applied to handwritten zip code recognition.Neural Comput. 1989; 1: 541-551Crossref Google Scholar). More layers and more neurons can represent more complex models; therefore, deep neural networks can be trained to classify complex objects, for example, images or videos. That increased complexity comes at the cost of increasing computations and computing power needed to train deep neural networks. A frequently used form of deep neural network is the convolutional neural network (CNN), in which groups of adjacent neurons from one layer feed into individual neurons in the next layer, thus extracting various features and capturing hierarchical patterns seen in images. CNNs are very powerful for image classification and object detection. Thus, they are widely used in biomedical imaging domains. With recent advancements in computing power and hardware design, particularly the availability of graphics processing units, efficient implementation of these deep neural networks is now feasible. Several deep learning frameworks and interfaces such as TensorFlow, Keras, and PyTorch support deep learning applications. TensorFlow, a Google product, is an open-source platform. It helps developers easily build machine learning models and neural networks. Keras is a high-level application programming interface of TensorFlow with a focus on modern deep neural networks. PyTorch is a different open-source library that was developed by Facebook’s AI Research Lab, and its Python interface is also popular for deep learning software development. Machine learning classifiers are assessed using several techniques. One is the receiver operating characteristics curve, which plots sensitivity (true positive rate) versus specificity (true negative rate) at different classifier output thresholds. Receiver operating characteristics curves provide a visual graphical way to assess a model’s performance and explain the trade-off between sensitivity and specificity. The area under the receiver operating characteristics curve is frequently used as an overall measure of classifier performance. This area under the curve (AUC) is an intuitive way of assessing the classifier. The AUC measures the performance of the classifier regardless of what classification threshold is used. The higher the AUC, the better job the classifier does at predicting. Most classifiers output the probability of belonging to a particular class. At a given probability threshold, sensitivity and specificity can also be reported. The accuracy, defined as the number of true predictions divided by the total number of predictions, is also frequently used. In binary classifications, the accuracy is the proportion of true positives and true negatives. Some of these values can be biased and misleading when class imbalance exists, that is, there are many more positive than negative examples or vice versa. Classifier performance needs to be estimated using data that were not used for training. Frequently, a fraction of data is withheld, for example, 10%–50%; the remaining data are used for training. Testing then occurs on the withheld data. In cross validation, the withheld set rotates through the data set and multiple models are trained and tested. As further developed later in this article, using the same data set for testing and training means the same biases exist in each subset. Thus, one or more fully independent replication test data sets are often needed to assess objective performances “in the real world.” Once shown to be accurate and predictive, machine learning models can be used to assess feature importance, that is, the contribution of individual features to the classifier’s accuracy. The recent development of deep neural networks has driven important AI and data science applications in medicine. For example, deep neural networks can be used to detect genetic variants (19Poplin R. Chang P.C. Alexander D. Schwartz S. Colthurst T. Ku A. et al.A universal SNP and small-indel variant caller using deep neural networks.Nat Biotechnol. 2018; 36: 983-987Crossref PubMed Scopus (274) Google Scholar) from large-scale genomic data sets and infer the functional impact of germline and somatic genetic variants (19Poplin R. Chang P.C. Alexander D. Schwartz S. Colthurst T. Ku A. et al.A universal SNP and small-indel variant caller using deep neural networks.Nat Biotechnol. 2018; 36: 983-987Crossref PubMed Scopus (274) Google Scholar, 20Zhou J. Park C.Y. Theesfeld C.L. Wong A.K. Yuan Y. Scheckel C. et al.Whole-genome deep-learning analysis identifies contribution of noncoding mutations to autism risk.Nat Genet. 2019; 51: 973-980Crossref PubMed Scopus (107) Google Scholar, 21Ainscough B.J. Barnell E.K. Ronning P. Campbell K.M. Wagner A.H. Fehniger T.A. et al.A deep learning approach to automate refinement of somatic variant calling from cancer sequencing data.Nat Genet. 2018; 50: 1735-1743Crossref PubMed Scopus (29) Google Scholar). Deep learning and other machine learning approaches may soon be used to predict responses to therapy based on genomic features (22Bhinder B. Elemento O. Towards a better cancer precision medicine: systems biology meets immunotherapy.Curr Opin Syst Biol. 2017; 2: 67-73Crossref PubMed Scopus (7) Google Scholar, 23Adam G. Rampášek L. Safikhani Z. Smirnov P. Haibe-Kains B. Goldenberg A. Machine learning approaches to drug response prediction: challenges and recent progress.NPJ Precis Oncol. 2020; 4Google Scholar). Deep neural networks have become instrumental in expanding precision medicine beyond genomics, especially in leveraging medical imaging to deliver more personalized medicine. A popular and powerful approach to using deep neural networks for image analysis involves using a technique called transfer learning, where parts of networks are pretrained on very large image data sets such as ImageNet, with more than 14 million images. ImageNet is currently a standard data set for pretrained networks. Such pretraining enables adaptation and fine-tuning on relatively small medical data sets, with perhaps only a few hundred images in each class (as opposed to a much larger data set required to train from scratch). Images and videos are perhaps where AI has been most transformative in medicine. For example, AI applied to skin lesion images can predict whether the lesion is malignant or not (24Esteva A. Kuprel B. Novoa R.A. Ko J. Swetter S.M. Blau H.M. et al.Dermatologist-level classification of skin cancer with deep neural networks.Nature. 2017; 542: 115-118Crossref PubMed Scopus (54) Google Scholar), and AI applied to retinal scans can predict diabetic retinopathies and other retinal diseases with high accuracy (25Gulshan V. Peng L. Coram M. Stumpe M.C. Wu D. Narayanaswamy A. et al.Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs.JAMA. 2016; 316: 2402-2410Crossref PubMed Scopus (3107) Google Scholar, 26De Fauw J. Ledsam J.R. Romera-Paredes B. Nikolov S. Tomasev N. Blackwell S. et al.Clinically applicable deep learning for diagnosis and referral in retinal disease.Nat Med. 2018; 24: 1342-1350Crossref PubMed Scopus (995) Google Scholar). Applied to pathology data (tissue images), AI can be used to differentiate between different cancer subtypes (27Khosravi P. Kazemi E. Imielinski M. Elemento O. Hajirasouliha I. Deep convolutional neural networks enable discrimination of heterogeneous digital pathology images.EBioMedicine. 2018; 27: 317-328Abstract Full Text Full Text PDF PubMed Scopus (148) Google Scholar), predict whether tumors have certain genetic alterations (28Coudray N. Ocampo P.S. Sakellaropoulos T. Narula N. Snuderl M. Fenyo D. et al.Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning.Nat Med. 2018; 24: 1559-1567Crossref PubMed Scopus (922) Google Scholar), diagnose disease from radiology images (29Lakhani P. Sundaram B. Deep learning at chest radiography: automated classification of pulmonary tuberculosis by using convolutional neural networks.Radiology. 2017; 284: 574-582Crossref PubMed Scopus (817) Google Scholar, 30Ardila D. Kiraly A.P. Bharadwaj S. Choi B. Reicher J.J. Peng L. et al.End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography.Nat Med. 2019; 25: 954-961Crossref PubMed Scopus (616) Google Scholar), and even identify polyps in colonoscopy videos (31Byrne M.F. Chapados N. Soudan F. Oertel C. Linares Perez M. Kelly R. et al.Real-time differentiation of adenomatous and hyperplastic diminutive colorectal polyps during analysis of unaltered videos of standard colonoscopy using a deep learning model.Gut. 2019; 68: 94-100Crossref PubMed Scopus (296) Google Scholar). Artificial intelligence applied to nonimaging data also shows promise. For example, deep learning has been applied to medical record data from thousands of patients from several centers and was shown to reliably predict the risk of readmission at 60 days, among other metrics (32Miotto R. Li L. Kidd B.A. Dudley J.T. Deep patient: an unsupervised representation to predict the future of patients from the electronic health records.Sci Rep. 2016; 6: 26094Crossref PubMed Scopus (738) Google Scholar, 33Rajkomar A. Oren E. Chen K. Dai A.M. Hajaj N. Hardt M. et al.Scalable and accurate deep learning with electronic health records.NPJ Digit Med. 2018; 1: 18Crossref PubMed Scopus (878) Google Scholar). Some of the developments in genomics have already led the broader field of reproduction and fertility treatment to embrace the philosophy of precision medicine. For example, carrier genetic screening can help parents make reproductive choices, such as whether preimplantation genetic testing should be used to select embryos that may not carry specific mutations. In older patients undergoing in vitro fertilization (IVF), preimplantation genetic testing for aneuploidy can help select embryos most likely to give rise to successful pregnancies. Preimplantation genetic testing may soon become less invasive if DNA from embryo cultures can be reliably sequenced and shown to mirror DNA in embryo cells (34Brouillet S. Martinez G. Coutton C. Hamamah S. Is cell-free DNA in spent embryo culture medium an alternative to embryo biopsy for preimplantation genetic testing? A systematic review.Reprod Biomed Online. 2020; 40: 779-796Abstract Full Text Full Text PDF PubMed Scopus (19) Google Scholar, 35Rubio C. Rienzi L. Navarro-Sanchez L. Cimadomo D. Garcia-Pascual C.M. Albricci L. et al.Embryonic cell-free DNA versus trophectoderm biopsy for aneuploidy testing: concordance rate and clinical implications.Fertil Steril. 2019; 112: 510-519Abstract Full Text Full Text PDF PubMed Scopus (44) Google Scholar, 36Vera-Rodriguez M. Diez-Juan A. Jimenez-Almazan J. Martinez S. Navarro R. Peinado V. et al.Origin and composition of cell-free DNA in spent medium from human embryo culture during preimplantation development.Hum Reprod. 2018; 33: 745-756Crossref PubMed Scopus (83) Google Scholar). Postimplantation genetic testing is also used to diagnose trisomy 21 and other genetic abnormalities. Likewise, noninvasive prenatal testing (liquid biopsy) in pregnant women can detect enough fetal DNA in the blood to diagnose genetic abnormalities, such as an abnormal number of chromosomes (37Fan H.C. Gu W. Wang J. Blumenfeld Y.J. El-Sayed Y.Y. Quake S.R. Non-invasive prenatal measurement of the fetal genome.Nature. 2012; 487: 320-324Crossref PubMed Scopus (289) Google Scholar) and genetic variants associated with Mendelian disorders (38Zhang J. Li J. Saucier J.B. Feng Y. Jiang Y. Sinson J. et al.Non-invasive prenatal sequencing for multiple Mendelian monogenic disorders using circulating cell-free fetal DNA.Nat Med. 2019; 25: 439-447Crossref PubMed Scopus (106) Google Scholar). Paradoxically, the same genomic technologies also reveal previously unknown complexity. For example, single-cell DNA sequencing has shown that 80% of embryos have mosaic aneuploidies (39Starostik M.R. Sosina O.A. McCoy R.C. Single-cell analysis of human embryos reveals diverse patterns of aneuploidy and mosaicism.Genome Res. 2020; https://doi.org/10.1101/2020.01.06.894287Crossref Google Scholar). Artificial intelligence has also begun to impact reproductive medicine, leading to further personalization (40Zaninovic N. Elemento O. Rosenwaks Z. Artificial intelligence: its applications in reproductive medicine and the assisted reproductive technologies.Fertil Steril. 2019; 112: 28-30Abstract Full Text Full Text PDF PubMed Scopus (16) Google Scholar). For example, deep learning can predict blastocyst quality based on static (41VerMilyea M. Hall J.M.M. Diakiw S.M. Johnston A. Nguyen T. Perugini D. et al.Development of an artificial intelligence–based assessment model for prediction of embryo viability using static images captured by optical light microscopy during IVF.Hum Reprod. 2020; 35: 770-784Crossref PubMed Scopus (13) Google Scholar) or time-lapse embryo images with high accuracy in individual patients (42Khosravi P. Kazemi E. Zhan Q. Malmsten J.E. Toschi M. Zisimopoulos P. et al.Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization.NPJ Digit Med. 2019; 2: 21Crossref PubMed Scopus (133) Google Scholar, 43Zaninovic N. Rosenwaks Z. Artificial intelligence in human in vitro fertilization and embryology.Fertil Steril. 2020; 114: 914-920Abstract Full Text Full Text PDF PubMed Scopus (30) Google Scholar). A CNN can also be trained to recognize specific areas in the embryo, such as the inner cell mass and the trophectoderm, which can then be fed into an algorithm that assesses embryo quality (44Kragh M.F. Rimestad J. Berntsen J. Karstoft H. Automatic grading of human blastocysts from time-lapse imaging.Comput Biol Med. 2019; 115103494Crossref Scopus (31) Google Scholar). As we have suggested, optimizing embryo selection may reduce the likelihood of multiple pregnancies and their associated risks. Deep learning can also be used to analyze sperm quality, thus helping to optimize intracytoplasmic sperm injection (45Riordon J. McCallum C. Sinton D. Deep learning for the classification of human sperm.Comput Biol Med. 2019; 111103342Crossref PubMed Scopus (39) Google Scholar, 46Iqbal I. Mustafa G. Ma J. Deep learning–based morphological classification of human sperm heads.Diagnostics (Basel). 2020; 10: 325Google Scholar). There are also many other applications of AI in obstetrics and gynecology, such as smarter fetal heart rate monitoring during pregnancy as well as prediction and detection of preterm labor and pregnancy complications (47Iftikhar P. Kuijpers M.V. Khayyat A. Iftikhar A. DeGouvia De Sa M. Artificial intelligence: a new paradigm in obstetrics and gynecology research and clinical practice.Cureus. 2020; 12: e7124Google Scholar). We anticipate that these exciting new applications of AI, such as individualization of hormone treatment, automated assessment of the uterus lining, and many others will continue to make reproductive medicine more precise and individualized, thus improving outcomes and limiting complications. Numerous challenges in the implementation of AI in medicine have been described (48He J. Baxter S.L. Xu J. Xu J. Zhou X. Zhang K. The practical implementation of artificial intelligence technologies in medicine.Nat Med. 2019; 25: 30-36Crossref PubMed Scopus (490) Google Scholar). Many of these apply to both precision medicine and AI, two fields that have reached maturity only recently. For example, precision medicine and AI both suffer from a relative lack of standardization. In genetic analysis, one can obtain results from a broad number of technologies, from microarrays to targeted capture panels to whole-genome sequencing. Comparisons between these platforms are limited, but those that have been done show some degree of divergence in the results (49Aguilera-Diaz A. Vazquez I. Ariceta B. Manu A. Blasco-Iturri Z. Palomino-Echeverria S. et al.Assessment of the clinical utility of four NGS panels in myeloid malignancies. Suggestions for NGS panel choice or design.PLoS One. 2020; 15e0227986Crossref Scopus (19) Google Scholar). This can likely be explained by varying capture performance between platforms and different variant calling and filtering thresholds. Likewise, in the AI field, there are numerous available techniques and software libraries for training AI models and a scarcity of large enough, publicly available data sets to robustly validate AI methods. A widely discussed challenge that applies to precision medicine driven by AI and genomics is the presence of biases in the data used to learn new medical knowledge or train predictive models (50Parikh R.B. Teeple S. Navathe A.S. Addressing bias in artificial intelligence in health care.JAMA. 2019; 322: 2377-2378Crossref PubMed Scopus (124) Google Scholar). To explain using a straightforward example, if a machine learning model is trained on two classes of image data, for example, good or bad images, and the number of bad images vastly outnumbers the good images, the AI predicting every image as bad will achieve good performance. In a less trivial example, data are collected from a specific demographic or from a particular type of imaging data. The classifiers trained on the data carry over the biases of the data and therefore may provide biased predictions when used prospectively. Genomic data sets that serve as the cornerstone of precision medicine have also been shown to have profound biases. The most visible one is the limited ethnic diversity present in the cohorts that these data sets come from (51Sirugo G. Williams S.M. Tishkoff S.A. The missing diversity in human genetic studies.Cell. 2019; 177: 26-31Abstract Full Text Full Text PDF PubMed Scopus (315) Google Scholar). The consequence of such biases is a precision medicine that may not only work better in some populations than others but may also lead to misdiagnoses in underrepresented populations (52Manrai A.K. Funke B.H. Rehm H.L. Olesen M.S. Maron B.A. Szolovits P. et al.Genetic misdiagnoses and the potential for health disparities.N Engl J Med. 2016; 375: 655-665Crossref PubMed Scopus (385) Google Scholar), thus potentially causing harm and exacerbating health disparities (53Martin A.R. Kanai M. Kamatani Y. Okada Y. Neale B.M. Daly M.J. Clinical use of current polygenic risk scores may exacerbate health disparities.Nat Genet. 2019; 51: 584-591Crossref PubMed Scopus (651) Google Scholar). It is critical that researchers who seek to learn from large clinical data sets and apply this knowledge to individual patients know how to look for such biases. They can then either try to limit them, for example, by pooling multiple data sets from multiple centers and sources, or to understand that what they learned; for example, the AI model they produced may not be applicable to certain data sets whose biases are distinct from the training data set. Whether it is precision medicine driven by AI, genomics, or both, replication across cohorts and prospective validation are also critical when learning of potential biases, which occurs when performances in replication cohorts are observed to be lower than in training cohorts. In a related issue, the size and representativeness of training data sets are a limitation in fields where the data are derived from patients. Some patient-derived data sets are publicly available, for example, the Human Sperm Head Morphology data set (HuSHeM) (54Shaker F. Human sperm head morphology dataset (HuSHeM).2018Google Scholar), but they tend to be small. Privacy safeguards such as HIPAA laws, the complexity of legacy electronic health record (EHR)/information technology (IT) systems, and competition between medical centers make it difficult to share data across centers and build large and diverse enough data sets for AI training. In certain fields such as cancer or radiology, some of these limitations have been overcome, and large public data sets have been built (55AACR Project GENIE ConsortiumAACR Project GENIE: powering precision medicine through an international consortium.Cancer Discov. 2017; 7: 818-831Crossref PubMed Scopus (625) Google Scholar, 56Weinstein J.N. Collisson E.A. Mills G.B. Shaw K.R. Ozenberger B.A. et al.Cancer Genome Atlas Research NetworkThe Cancer Genome Atlas Pan-Cancer analysis project.Nat Genet. 2013; 45: 1113-1120Crossref PubMed Scopus (3781) Google Scholar). Novel forms of machine learning such as federated learning may one day help address some of these data-sharing issues (57Yang Q. Liu Y. Chen T. Tong Y. Federated machine learning: concept and applications.ACM Trans Intellig Syst Techn. 2019; 10Google Scholar). Implementation is also a challenge. This is because AI software needs to integrate into existing, clinically validated workflows that need revalidation and therefore constitute a barrier to implementation. Because most AI is based on pattern matching and does not imply any reasoning, it does not manage borderline cases as well as a reasoning, medically trained human. One approach toward integrating AI and human interaction is via human machine collaboration, an active area of investigation (58Tschandl P. Rinner C. Apalla Z. Argenziano G. Codella N. Halpern A. et al.Human-computer collaboration for skin cancer recognition.Nat Med. 2020; 26: 1229-1234Crossref Scopus (153) Google Scholar, 59Patel B.N. Rosenberg L. Willcox G. Baltaxe D. Lyons M. Irvin J. et al.Human-machine partnership with artificial intelligence for chest radiograph diagnosis.NPJ Digit Med. 2019; 2: 111Crossref PubMed Scopus (50) Google Scholar). For skin cancer diagnosis, AI-based decision support was shown to improve diagnostic accuracy compared with either AI or physicians alone; moreover, it appeared that less experienced clinicians may benefit more from AI-based support than more experienced ones (58Tschandl P. Rinner C. Apalla Z. Argenziano G. Codella N. Halpern A. et al.Human-computer collaboration for skin cancer recognition.Nat Med. 2020; 26: 1229-1234Crossref Scopus (153) Google Scholar). Likewise, the diagnosis of pneumonia on chest radiographs was improved when groups of radiologists and deep learning algorithms were combined compared with radiologists or deep learning alone (59Patel B.N. Rosenberg L. Willcox G. Baltaxe D. Lyons M. Irvin J. et al.Human-machine partnership with artificial intelligence for chest radiograph diagnosis.NPJ Digit Med. 2019; 2: 111Crossref PubMed Scopus (50) Google Scholar). Such collaborations between humans and AI may become frequent in the future. For example, AI may be able to make largely automated decisions in most cases, but some cases may need complex reasoning that cannot yet be learned by AI. Precision medicine, a form of medicine that seeks to individualize treatment beyond what is currently done, is attracting substantial interest. Artificial intelligence and data science, two fields that have only recently achieved maturity, will increasingly play a core role in expanding the reach of precision medicine. Such evolution will encounter significant challenges. In this review, we sought to present an overview of precision medicine and AI. In particular, we sought to introduce basic concepts but also challenges and limitations in AI that span many if not most of the impactful AI applications to date. Artificial intelligence is already impacting many medical fields, and we predict that it will impact reproductive medicine in a major way. We envision that both precision medicine and AI will play a key role in the IVF clinic of the future, improving outcomes but also reducing pregnancy complications and allowing couples to be in better control of their reproductive process.