Exploring the ability of all features, clinical, TLT and patient’s characteristics to distinguish BL versus nBL embryos, become clear that dimensionality reduction methods, as well as Principal Component Analysis, ineffective in discerning between the embryos.
In the present study, we describe the performance of the novel EmbryoMLSelection framework in identifying a set of rules associated to a timely embryo development to the expanded blastocyst stage on day 5, called
feature-signature. The rationale behind a two-step process of features selection, rule extraction and rule selection from a large number of variables/embryo (
n = 30) was the identification of a set of rules (from an initial number of 71 to the final 6) identifying an
feature-signature composed of relevant features (
n = 8), describing cleavage stage embryos able to timely (within day 5) progress to the expanded blastocyst stage. As lower implantation and clinical pregnancy rates were reported in case of transfer of slow-growing blastocysts vs. fully expanded day 5 blastocysts [
34,
35], probably due both to a poor embryo competence and to the loss of embryonic-endometrial synchrony [
36], we considered the fully expanded blastocyst on day 5 as the optimal development stage, that confers the highest probability of embryo implantation. In fact, the number of days of blastocyst development represents the developmental potential of a blastocyst [
37] and affects the outcome of transfer [
38]. In addition, the developmental potential of Day 5, 6 and 7 blastocysts decreases gradually with the extension of culture time [
39]. Therefore, the conventional practice in the laboratory is to select blastocysts for transfer, biopsy or cryopreservation, starting from expanded day 5 blastocysts. In our dataset, 36.5% progressed to the expanded blastocyst stage on day 5 (BL group defined according to the score 3 provided by the Istanbul Consensus) whereas 63.5% did not (nBL Group). This data should be discussed in relation to the overall blastocyst formation rate of 53.8% observed in the enrolled patients. In fact, according to the Vienna Consensus [
40], we considered also blastocysts with expansion score of 2 in the KPI “blastocyst formation rate”, excluded from the BL group of our study. For this reason, we believe we can exclude any selection bias from the patients population ensuring an adequately powered analysis during the framework development. Embryo selection models developed using morphokinetic parameters were previously shown to predict blastocyst development [
41,
42]. Furthermore, the application of machine-learning technology provided an algorithm able to predict clinical pregnancy and live birth rate by analysing embryo morphokinetics [
43]. Giscard d’Estaing [
18] used a machine-learning system in order to build up a score for blastocyst formation with a prediction power having AUC = 0.634. In another study, the prediction accuracy of embryo assessment performed by experienced embryologists with morphokinetic grading methods added to conventional static morphology was shown to range between 60% and 70%, with AUC = 0.63–0.70 [
44]. In a previous study, the efficacy of six in-house embryo-selection algorithms (ESAs) was investigated in a set of known implantation embryos [
45]. Interestingly, although the primary endpoint considered and the nomenclature adopted were different, we both observed that tPNf, s2 (i.e. t4-t3) and cc3 (i.e. t8-t4) were associated to embryo developmental and reproductive competence. Since their results highlighted that ESAs are usually specific to the patient, treatment, and environment, we agree that currently available algorithms should be carefully validated before consider clinical applicability and, more importantly, they risk to lose their diagnostic value when externally applied. Herein, we provide evidence that the novel EmbryoMLSelection framework allowed to perform a more precise evaluation of embryo dynamic growth with a performance described by AUC = 0.84 and accuracy of 81%. Notably, the Rules Selection step ensured such an increased performance providing a concomitant reduction of the rules and variables used (
n = 6 and 8, respectively). Importantly, the EmbryoMLSelection framework developed here was registered in the Docker image and therefore its application is globally accessible online. Of note, the rules associated with the ability of reaching the stage of expanded blastocyst on day 5 include early embryo-related variables, such as embryo morphological score on day 2, and some cytokinesis times occurring in the first three days of development (tPNf, t4, t4-t3, t8-t4). So far, only one study coupled TLT annotation with morphological embryo assessment performed with the evidence-based score named IMCS [
46]. According to our results, good quality embryos having static morphological score > 6.0 on day 2 are more likely to reach the expanded blastocyst stage on day 5. In addition, the relevance of timings describing early embryo development is confirmed by previous studies reporting that a timely blastocyst development on day 5 can be predicted looking at the first three days of development [
28,
47]. Moreover, morphokinetic data of cleavage stage embryos were found to be associated to both blastulation rate and blastocyst quality [
48]. Indeed, embryos with quicker cleavage time from the 2-cells to the 8-cells stage have the highest potential to timely become blastocyst with good morphological score, and with the ability to expand and implant [
49,
50]. In this context, the pivotal clinical significance of our framework would be to indicate on day 3 which embryos are more likely to develop into viable blastocysts, giving the potential advantage to select the most competent embryos on day 3 without the need to extend culture till day 5, thus saving time and resources.
Moreover, other clinical variables, such as age, AFC and OSI, were associated to the timely progression to the blastocysts stage.
Indeed, female age defined as advanced (AMA; >35 year) was extensively associated with a decline in oocyte yield, fertilization, and overall oocyte/embryo developmental competence, mainly due to an increased incidence of aneuploidies and a decreased mitochondrial activity [
51,
52]. Studies reporting embryo morphokinetics from the fertilization to the pre-implantation period in women of AMA remain limited; in our dataset, 44 (55%) of cycles were performed in patients with AMA and a total of 303 (53%) embryos were included in our analysis suggesting a link between morphokinetic pattern and maternal age. Maternal age seems to have a relevant impact on the regulation of cell polarity during compaction, as well as on blastocoel cavity expansion, suggesting that AMA may affect embryo competence irrespective of the well-known consequences of oocyte meiotic errors [
53]. On the other hand, AFC and OSI are markers of ovarian reserve and responsiveness to COS, and are associated not only with female age, but also with circulating AMH levels, oocyte yield and, ultimately, clinical pregnancy [
54,
55].
Importantly, the rules proposed in this study are presented as a signature rather than a machine learning model. Despite the AdaBoosting algorithm had the highest AUC value in discriminating between BL and nBL and provided a model comprising rule-to-weight associations which influence each rule’s contribution to the final prediction, our focus was primarily to identify the rules rather than their weight.