Table 3 Trigger day assessment studies using artificial intelligence
From: The prospect of artificial intelligence to personalize assisted reproductive technology
Study | Aims of study | Outcomes of interest | Dataset | AI methods | Results |
|---|---|---|---|---|---|
Abbara et al. (2018) 25 | Follicle sizes on the day of trigger most likely to yield mature oocytes. | Optimal follicle sizes on TD. | Retrospective dataset with 499 patients. GnRH-ant protocol with hCG (n = 161); GnRH-a (n = 165); KP-54 (n = 173) triggers. Input variables: individual follicle diameters (in mm) from ultrasound scan on TD. | Random forest with 5-fold CV. | Follicles of diameter 12–19 mm were most contributory to the models following all three trigger types. |
Abbara et al. (2020)19 | Examine the relationship between endocrine changes following the use of different oocyte maturation triggers. Assess the relative importance of endocrine predictors when predicting mature oocyte yield. | (1) Accuracy in predicting the number of mature oocytes retrieved. (2) Relative importance of LH/hCG as an input variable. | Retrospective dataset with 499 patients. GnRH-ant protocol with hCG (n = 161); GnRH-a (n = 165); KP-54 (n = 173) triggers. Input variables: baseline endocrine characteristics, number of follicles sized 12–19 mm. | Performance comparison between a random forest with 5-fold CV and an ANN model. | Random forest had 88% accuracy within a tolerance level of 3 mature oocytes. The performance dropped to 83% when data on baseline LH/hCG levels were excluded. -The ANN had 57% accuracy. |
Robertson et al. (2021)32 | Finding the optimal tracking strategy for OS to minimize face-to-face interactions. | Earliest day during OS which can predict both the optimal TD and risk of OHSS accurately. | Retrospective dataset with 2128 cycles of 1731 women in a single center. 88.8% were GnRH-ant (fixed) cycles. An hCG trigger was used. Input variables: age, AFC, follicle count by size on each scan. | Random forest regressor for TD. Binary random forest classifier for OHSS prediction. | Day-5 was the earliest cycle day for predicting both outcomes accurately. The day-5 model had a MSE of 2.16 ± 0.12 for TD and AUC of 0.91 ± 0.01 for OHSS classification. |
Hariton et al. (2021)26 | Optimize TD timing to maximize 2PN and usable blastocyst yield. | Average improvement of 2PNs (primary outcome), and usable blastocysts vs. a clinician’s decision. | Retrospective dataset with 7,866 ICSI cycles. 1,967 cycles (25%) held out for independent testing. GnRH-ant, LD21, Lupron stop, flare, or mini-IVF (natural cycle) protocols were used. Input variables: age, BMI, number of follicles of {6-10, 11-15, 16-20, 21-25} mm, E2 level, protocol type, TD. | Light Gradient Boosting Machine with bagging. | Average improvement: 3.015 more 2PNs (95% CI 2.626, 3.371) and 1.515 more usable blastocysts (95% CI 1.134,1.871). Given physician agreement with the model (52.57% for 2PNs, 61.89% for blastocysts): 1.430 more 2PNs, and 0.577 more usable blastocysts. Follicle sizes 16-20mm were most contributory to the model performance. |
Letterie et al. (2022)31 | Workflow optimization of OS: (1) single ‘best day’ for monitoring during OS; (2) predict optimal TD; (3) predict total number of retrieved oocytes. | Acc., TPR, and PPV of total number of retrieved oocytes and mature oocytes stratified into: 0–10, and >10. MAE of determining the aims of (1) and (3). | Retrospective data-set with 1591 IVF cycles from a single center. 318 cycles (20%) held out for independent testing. An hCG or Lupron trigger was used. Pre-cycle selected input variables: age, AMH. ‘Best day’ selected input variables: E2 levels, follicle counts and sizes, day of cycle during OS, dose of FSH during OS. | Stacking ensemble model comprising: linear regression, random forest, extra trees regression, k-nearest neighbors, XGBoost. | (1) ‘Best day’ prediction: MAE 1.355. (2) Variance of 0-3 days for trigger choice showed “little impact” to oocytes retrieved. (3) Total number of oocytes: MAE 3.517. Total retrieved oocytes: Acc. 0.77; 0-10 oocytes (TPR 0.80; PPV 0.79); >10 oocytes (TPR 0.74; PPV 0.74). Total retrieved mature oocytes: Acc. 0.89; 0-10 oocytes (0.91; 0.89); >10 oocytes (0.86; 0.88). Total number of oocytes: MAE 3.517. |
Fanton et al. (2022)27 | Optimize TD timing to maximize MIIs, 2PNs, and blastocyst yield. | Average number of MIIs (primary outcome), 2PNs, and usable blastocysts. | Retrospective dataset with 30,278 cycles from 3 centers (2555, 3051, 14,672 cycles). 20% held out for independent testing. No available protocols were excluded. Pre-cycle input variables: age, BMI, AFC, previous IVF cycles, AMH, E2 level, cycle length (in days). Mid-cycle input variables: number of follicle scans during OS, E2 levels during OS, number of follicles of size {<11, 11–13, 14–15, 16–17, 18–19, >19} mm on TD. | Multivariable linear regression | Patients with early triggers had 2.3 fewer MIIs, 1.8 fewer 2PNs, and 1.0 fewer usable blastocysts when compared to propensity-matched on-time triggers. Patients with late triggers had 2.7 fewer MIIs, 2.0 fewer 2PNs, and 0.7 fewer usable blastocysts when compared to propensity-matched on-time triggers. Only follicle sizes and E2 were used in the final model. |
