CA | ES | EN
AI with care: Integrating machine learning with expert knowledge for In Vitro Fertilization
AI with care: Integrating machine learning with expert knowledge for In Vitro Fertilization
Núria
Núria
 
Correa Mañas
Correa Mañas
 (
24/Jul/2023
24/Jul/2023
)
AI with care: Integrating machine learning with expert knowledge for In Vitro Fertilization
AI with care: Integrating machine learning with expert knowledge for In Vitro Fertilization
 

An industrial PhD

Advisors: 

Josep Lluís Arcos Rosell

Rita Vassena

Jesús Cerquides Bueno

Josep Lluís Arcos Rosell

Rita Vassena

Jesús Cerquides Bueno

University: 

Abstract: 

After almost 45 years after the birth of Lousie Brown, the first baby born after in vitro fertilization (IVF), pregnancy rates for this treatment remain around 30%, with a 20% chance of delivery. Even if it is much better than the chances that those patients had without IVF, logically, there are constant endeavors to gain insight into the biological reality behind fertility in order to refine artificial reproduction technologies (ART).

In parallel to the technical advances achieved by ART professionals, artificial intelligence (AI) has also progressed at a remarkable pace. Its ability to deal with high dimensional databases and detect hidden data relationships has led researchers to investigate its application to healthcare. There are several processes in ART, and specifically, IVF, where AI methods are currently being applied.

In this thesis, the main focus is on the selection of the first dose of follicle-stimulating hormone (FSH) for controlled ovarian hyperstimulation (COH). COH is the first step of an IVF treatment, where the objective is to retrieve an optimal number of mature oocytes from the ovary. Its results are critical for the success of the IVF treatment. Standard clinical protocols to select the first dose of FSH are not perfect and lead a sizable portion of patients to suboptimal results. Here, we use AI methods with historical data from past COH treatments to obtain an optimized FSH dosing policy.

Historical or observational datasets are often biased and prone to low variability due to the high adherence of clinicians to standard protocols. In this context, out-of-the-box AI methods do not have enough information to learn dosing models that improve standard practice, or even show consistency with the underlying physiological reality. Hence, the introduction of domain knowledge in the training process is key to obtaining clinically robust models from observational data. To achieve this, we propose building the dosing model around the assumption that the dose-response relationship between FSH and the number of oocytes retrieved is monotonic.

Further, since insight into the performance of dosing models is generally achieved through prospective clinical intervention, we designed an ad-hoc performance score to evaluate doses (real or counterfactual) pre-clinically. This score, based on expert knowledge, can evaluate whether a dose is appropriate depending on the ground truth outcome, expressed as the number of mature oocytes retrieved. Using this method, we were able to ascertain a statistically significant improvement versus standard clinical practice. A generalized method for similar dosing problems, called IDoser, was also tested in the FSH use case against clinical practice and a benchmark of literature, finding again significant improvement. A first approach of the application of IDoser to the selection of the number of embryos for transfer in IVF also returned positive results with potential for improvement.

Finally, AI-driven solutions, especially for healthcare settings like drug dose selection, are to be handled with care, as patients’ health is at stake. Not only that, they need to gain the trust of their intended users, in this case, clinicians. Trust is gained through clinical improvement and easy-to-prove adherence to already available field knowledge. The first can be achieved through pre-clinical analysis, but especially through randomized controlled trials (RCTs) where the models are tested against standard practice. The second, as proposed in this thesis, can be achieved through interpretable implementations of domain knowledge in the creation and training of dosing models. This leads to clinically robust dosing models that achieve better pre-clinical results.

After almost 45 years after the birth of Lousie Brown, the first baby born after in vitro fertilization (IVF), pregnancy rates for this treatment remain around 30%, with a 20% chance of delivery. Even if it is much better than the chances that those patients had without IVF, logically, there are constant endeavors to gain insight into the biological reality behind fertility in order to refine artificial reproduction technologies (ART).

In parallel to the technical advances achieved by ART professionals, artificial intelligence (AI) has also progressed at a remarkable pace. Its ability to deal with high dimensional databases and detect hidden data relationships has led researchers to investigate its application to healthcare. There are several processes in ART, and specifically, IVF, where AI methods are currently being applied.

In this thesis, the main focus is on the selection of the first dose of follicle-stimulating hormone (FSH) for controlled ovarian hyperstimulation (COH). COH is the first step of an IVF treatment, where the objective is to retrieve an optimal number of mature oocytes from the ovary. Its results are critical for the success of the IVF treatment. Standard clinical protocols to select the first dose of FSH are not perfect and lead a sizable portion of patients to suboptimal results. Here, we use AI methods with historical data from past COH treatments to obtain an optimized FSH dosing policy.

Historical or observational datasets are often biased and prone to low variability due to the high adherence of clinicians to standard protocols. In this context, out-of-the-box AI methods do not have enough information to learn dosing models that improve standard practice, or even show consistency with the underlying physiological reality. Hence, the introduction of domain knowledge in the training process is key to obtaining clinically robust models from observational data. To achieve this, we propose building the dosing model around the assumption that the dose-response relationship between FSH and the number of oocytes retrieved is monotonic.

Further, since insight into the performance of dosing models is generally achieved through prospective clinical intervention, we designed an ad-hoc performance score to evaluate doses (real or counterfactual) pre-clinically. This score, based on expert knowledge, can evaluate whether a dose is appropriate depending on the ground truth outcome, expressed as the number of mature oocytes retrieved. Using this method, we were able to ascertain a statistically significant improvement versus standard clinical practice. A generalized method for similar dosing problems, called IDoser, was also tested in the FSH use case against clinical practice and a benchmark of literature, finding again significant improvement. A first approach of the application of IDoser to the selection of the number of embryos for transfer in IVF also returned positive results with potential for improvement.

Finally, AI-driven solutions, especially for healthcare settings like drug dose selection, are to be handled with care, as patients’ health is at stake. Not only that, they need to gain the trust of their intended users, in this case, clinicians. Trust is gained through clinical improvement and easy-to-prove adherence to already available field knowledge. The first can be achieved through pre-clinical analysis, but especially through randomized controlled trials (RCTs) where the models are tested against standard practice. The second, as proposed in this thesis, can be achieved through interpretable implementations of domain knowledge in the creation and training of dosing models. This leads to clinically robust dosing models that achieve better pre-clinical results.