Thread by @juanjo_fraire

And here we are again for February #ESHREjc 🤓🤓 This months we are discussing “Predictive models of pregnancy based on data from a preconception cohort study” from @jennifer_yland @YPaschalidis @LaurenAnneWise
@ESHRE

doi.org/10.1093/humrep…

We wanted to transform our #ESHREjc into a forum where clinicians/embryologists get interested into what #MachineLearning is and how it can be used in #reproductivehealth or #ReproductiveMedicine
For an intro please see the following summary 👇🏼👇🏼👇🏼

https://twitter.com/juanjo_fraire/status/1495409035994148867

Are we ready? Let’s dive into the article #ESHREjc

Yland et al, ask: “Can we derive adequate models to predict the probability of conception among couples actively trying to conceive?” #ESHREjc

1.1 INTRO
In North America, 10-15% of couples experience infertility. In the USA alone, 12% of reproductive aged women (25-44 y.o.) and 9% of men used fertility treatments from 2006 to 2010 costing 5 billion USD 💸😅 #ESHREjc

1.2 INTRO
So, better prognostic tools would help to mitigate potential costs and inform couples about their fertility potential before they start trying to conceive #ESHREjc

2.1 METHODS
All right! where to start? 🤔
Age, BMI are risk factors for infertility. Preconception exposures (alcohol, smoking, stress, diet, sleep quality, etc) are related to reduced fecundability 😳 however few studies with modest predictive power! #ESHREjc

2.2 METHODS
@jennifer_yland et al, used Supervised Machine Learning (ML) to predict:

a) Cumulative probability of pregnancy (over 6 and 12 menstrual cycles)

b) Fecundability (per-cycle probability of conception)

#ESHREjc

2.3 METHODS
How? 🧐

➡️ 163 potential predictors (variables)
➡️ several classification algorithms and variable selection procedures

🤓 to identify the most accurate models and evaluate relative predictive strength of individual risk factors #ESHREjc

2.4 METHODS
#ESHREjc
Study population using the Pregnancy Study Online (PRESTO) (Wise et al, 2015)
A web-based cohort study of couples in North America.
➡️ 21-45 years old females
➡️NOT using contraception or fertility treatments

➡️Data collected on ≠ factors:
-Sociodemographic
-Behavioral
-Medical/Reproductive history
-Male partner characteristics

➡️ Diet History Questionnaire (DHQ)

➡️ Bimonthly follow-up questionnaires until:
-pregnancy
-cessation of pregnancy attempts
-study withdrawals
#ESHREjc

2.5 METHODS
Three models to predict:

I) pregnancy in <12 menstrual cycles (infertility)
N= 3195

II) pregnancy within 6 menstrual cycles (subfertility)
N=3476

III) average probability of pregnancy per menstrual cycle (fecundability)
N=4133

#ESHREjc

2.6 METHODS
#ESHREjc
👉🏼Data were pre-processed for feature selection and avoid overfitting 🤓🦾
👉🏼Dataset randomly split in training and testing sets
👉🏼Algorithm applied to infer input (predictor) to outputs (pregnancy outcome)

#ESHREjc
👉🏼Statistical feature selection to improve prediction of outcome

👉🏼 Then, compare 4 supervised classification methods for pregnancy predictive models:

#ESHREjc
For Model I (infertility) and II (subfertility)

➡️Logistic regression models with regularization (robust to outliers)

➡️Support Vector Machines (SVM) also with regularization

➡️Light Gradient Boosting Machine (LightGBM)

➡️Artificial Neural Network

For Model III (fecundability)
➡️discrete-time analog of Cox proportional hazards models 😱

3.1 RESULTS
😅 #ESHREjc

~30 years old study participants
➡️86% participants in Model I became pregnant (<12 mens cycles)
➡️69% in Model II (6 mens cycles)

After a PARSIMONIUS model of feature selection (good fit using few explanatory variables) 👇🏼

- 14 variables for Model I (infertility)
- 15 variables for Model II (subfertility)

Performance of model was evaluated using Area Under the ROC Curve (AUC) where 0.5 is no better than random and 1.0 is perfect prediction

3.2 RESULTS
#ESHREjc
🦾 68-70% AUC for Model I (infertility; <12 mens cycles) for all algorithms.

🦾 65-66%AUC for Model II (subfertility; 6 mens cycles)

So good performance? 🤔
Previously published models had AUCs from 59% to 63%

#ESHREjc
Regularized models (robust for outliers) for many variables (L2) in both Logistic Regression (L2LR) and SMV (L2SVM) yielded the highest AUC 🦾

Concordance index for Model III (fecundability) was 62.6% for final parsimonious model
#ESHREjc

3.3 RESULTS
#ESHREjc
Variables in parsimonious Model I (infertility) POSITIVELY associated with pregnancy:

-Menstrual cycle length
-Living in rural region
-Use of vitamins or folic acid
-Use of IUD
-previously breastfed an 👶
-been 🤰🏻
-Education
-flu 💉
-Gravidity

Variables in parsimonious Model I (infertility) INVERSELY associated with pregnancy:

-👩‍🦳 age
-History of infertility
-Having completed 1 menstrual cycle of 🤰🏻 attempt time at study entry
-👩‍🦳 BMI
-Stress
#ESHREjc

#ESHREjc
Variables in parsimonious Model II (subfertility) POSITIVELY associated with pregnancy:

-Use of vitamins or folic acid
-Previously breastfed an 👶
-Diet quality
-Previous unplanned 🤰🏻
-Improving chances of 🤰🏻 (intercourse in fertile window)
-Time since last 🤰🏻

Variables in parsimonious Model II (subfertility) INVERSELY associated with pregnancy:

-👩‍🦳 BMI
-History of infert
-👨‍🦰 age
-non-use of fert app
-👨‍🦰 BMI
-Having completed 1 mens cycle of 🤰🏻 attempt
-👨‍🦰 partner 🚬
-👨‍🦰 age
-History of sub/infertility
#ESHREjc

Variables selected into Model III (fecundability) but neither Models I or II included:

-Intercourse frequency
-Menstrual cycle regularity
#ESHREjc

And now a twist 😉 #ESHREjc

3.4 RESULTS
#ESHREjc
Among 1957 NULLIGRAVID women w/out a history of infertility, models were developed for predicting pregnancy for infertility (Model IV), subfertility (Model V) and fecundability (Model VI).

Performance slightly lower than full cohort!

#ESHREjc
5 variables selected for Infertility Model and 9 variables selected for Subfertility Model.

L2 Logistic Regression algorithm performed best 🤓

🦾69.5% AUC Infertility Model
🦾65.6% AUC Subfertility Model

Concordance index for Fecundability Model was 60.2%

Variables POSITIVELY associated with 🤰🏻
#ESHREjc
-menstrual cycle length
-hormonal IUD
-intercourse frequency
-improving chances of 🤰🏻
-use vitamin E supplements
-diet quality

Variables INVERSELY associated with 🤰🏻

-completed 1 mens cycle of 🤰🏻 attempt
-👩‍🦳 age
-👨‍🦰and 👩‍🦳 BMI
-menstrual cycle irregularity
-non-use of fert app
-stress
-depressive symptoms
-history of vaginosis
-male partner 🚬
-milk consumption
- 😴 characteristics

4.1 CONCLUSIONS

➡️ Authors claim their study may be generalizable to couples across fertility spectrum

➡️Findings relevant for couples planning a pregnancy

➡️If models validated in external populations, potential counseling tool
#ESHREjc

5.1 LIMITATIONS

❎ Self-reported predictor data could have introduced misclassifications

❎ Not certain all relevant predictor variables were considered

Let’s start the discussion with authors @jennifer_yland @YPaschalidis @LaurenAnneWise and experts @DrMPerugini and @vlbthambawita 🤓🦾
#ESHREjc @ESHRE