FIELD OF THE INVENTION
This invention relates to the field of healthcare and particularly relates to an artificial intelligence assisted health
activist and pregnant women.
BACKGROUND OF THE INVENTION
Artificial intelligence (AI) studies 'agents' that respond to their surroundings through activities. These are called
'intelligent agents' AI refers to how computers can execute human-like jobs. This includes activities like
document translation, face recognition, and decision-making. In this review, we will focus on clinical and patient
decision-making, as AI can alter pharmacological or medication decisions during pregnancy. Initially, healthcare
AI used rule-based decision-making. AI systems that use rule-based decision-making fit easily in clinical
environments because they reflect physicians' decision-making. MYCIN was a rule-based decision-making
algorithm. MYCIN was discovered in 1974 to forecast bacterial infection therapies. It was developed to guide
clinicians to proper decisions using if–then sentences. In 1994, 'expert' rule-based systems were first applied to
women's health with the invention of a pre-term birth risk predictor. This AI-based rule-based method predicted
a woman's probability of preterm birth using diagnostic codes during pregnancy. Restructuring improves these
programmes' performance. ML is an AI application that permits unprogrammed learning. An artificial neural
network (ANN) mimics how biological brain networks process data.
SUMMARY OF THE INVENTION
This demonstrates how AI has been applied to address pharmacological exposures during pregnancy,
and this includes the entire pregnancy process: preconception, prenatal, prenatal, and postnatal health
concerns. We identify three areas where AI methods could be used to improve our understanding of the
pharmacological effects of pregnancy, including obtaining sound and reliable data from clinical records
and designing optimized animal experiments to validate specific hypotheses by implementing decision
support systems that inform decision-making. The largest literature gap that we identified is with regard
to using AI methods to optimize translational studies between animals and humans for pregnancy-related
drug exposures. However, in general, all three areas lacked research regarding the pharmacological
exposure-pregnancy aspect with less than per category. Incorporating modern AI methods into
understanding the maternal and fatal consequences of pharmacological drug exposure is a must for
future studies. Applications of AI to other aspects of pregnancy, maternal, and fetal health, including
lactation, can inform the necessary research to delve more deeply into how pharmacy logic affects
pregnancy.

BRIEF DESCRIPTION OF THE INVENTION
The field of artificial intelligence (AI) involves the study of ‘agents’ that receive information from their environment and perform actions in response to that environment. These ‘agents’ are sometimes referred to as ‘intelligent agents. In general, AI is used to refer to the method by which computer systems can perform tasks that would typically require a human. This includes tasks such as translating documents into different languages, automatically identifying a person from an image, visual perception or decision-making. In this review, we will focus our discussion on both clinical and patient decision-making, as these are two areas where AI has the potential to impact decision-making with respect to pharmacological or drug choices during pregnancy. Initially, AI methods that were used in healthcare focused on rule-based decision-making. AI tools that utilize rule-based decision-making fit naturally within the clinical environment because they can effectively mirror the clinicians’ own decision-making process. One of the first rule-based decision-making algorithms was MYCIN. MYCIN was developed in 1974 to predict the appropriate therapy for different bacterial infections. It was designed as an ‘expert system’ that would guide clinicians to appropriate decision-making, using a series of if-then statements. These ‘expert’ rule-based systems would be first applied in the field of women’s health some 20 years later in 1994 with the development of a rule-based preterm birth risk predictor. This rule-based system predicted a woman’s risk of preterm birth using diagnostic codes during the pregnancy and utilized the ‘state-of-the-art’ in AI at that time. These types of programs can only achieve improved performance through restructuring. Machine learning (ML) is an application of AI that enables learning without being explicitly programmed. A popular method of ML, an artificial neural network (ANN), is designed to resemble how biological neural systems process data.
AI is the broad science of mimicking human abilities. Machine learning is a subset of AI in the field of computer science. ML often uses statistical techniques to allow for the computer to ‘‘learn’’ or progressively improve performance on a given task without being explicitly programmed. ML refers to a number of methods and algorithms and different learning types: supervised, semi-supervised, unsupervised, reinforcement, evolutionary, and deep learning. In supervised learning, every input pattern is trained to an associated output pattern, and errors in computed and desired outputs can be used to improve performance. Common supervised learning algorithms include regression and classification algorithms, such as the following: simple linear regression, polynomial regression, LASSO regression, k-Nearest Neighbors, Support Vector Machines (SVM), Na¨ive Bayes (NB), Decisions Trees (DT), and Random Forests (RF). In unsupervised learning, the network trains without the knowledge of the desired output. Common unsupervised learning methods include clustering algorithms and dimension reduction algorithms, such as k-means clustering, principal component analysis (PCA), and independent component analysis (ICA). In reinforcement learning, agents are not presented, and the desired output is learned from the actions that are the best through trial and error. ML models learn from a given dataset with instances and features; an instance is an individual or example in the data. Each instance has a number of features or attributes describing an aspect of that instance. Important care needs to be taken when considering different ML techniques for a classification problem. How- ever, this decision is situational and dependent on the dimensionality, size, and other qualities of the dataset. ML methods are not designed to demonstrate causality and, at best, can provide likely candidates for causality. No single model performs optimally across all problems, and this phenomenon is called the No Free Lunch theorem. For this reason, it is common to compare more than one modelling approach, compare models with different parameters, or develop an ensemble approach. For the sake of this review, we will not delve deeply into the advantages and disadvantages of AI methods. A recent perspective article provides an overview of the barriers to deployment and the translational impact of ML methods for health care. The operation and fitting of ML methods, the ethics of AI in medicine, as well as unintended consequences are comprehensively discussed elsewhere.
BETTER ‘ACTIONABLE’ SCIENCE NEEDED FOR PREGNANT WOMEN
Twenty-five years have passed since the first AI tool was developed for a woman’s health issue (preterm birth) and a full 45 years since the first health-based ‘expert’ AI system was developed. However, much remains to be done in the realm of harnessing AI methods to improve healthcare, especially women’s health. Recently, two articles in the New England Journal of Medicine highlighted the important need for novel methods to investigate pharmacological effects among pregnant women and also nursing or lactating women. Both Eke et al. and Mitchell et al. could be used to design optimal animal models for experiments that validate retrospective findings obtained from clinical records or other sources. c illustrates how artificial intelligence or machine learning methods could be used to alert physicians and their patients at the appropriate time pertaining to specific details related to pregnant or nursing women note that 70–80% of pregnant women receive a pharma- cologic during their first trimester of pregnancy (the most critical period with regards to congenital anomalies and adverse fetal outcomes) and 90% of pregnant women take a pharmacologic at any point during their pregnancy. Moreover, prescription medication use increases with maternal age and education . However, the majority of these medications are prescribed and taken without any randomized controlled trials that include pregnant women and this limits the ability for clinicians to understand the potential adverse health consequences both for the mother and the fetus. A tremendous need exists to understand the effects of pharmacologics not only on the developing fetuses in terms of anomalies and other adverse fetal consequences, but also with regards to the mother. Adverse maternal consequences are possible and include excessive bleeding and other perinatal and postpartum complications. Pregnant and postpartum women have been systematically excluded from research due to their vulnerable status.
Pregnant women may be viewed as scientifically complex, and there are practical and ethical issues sur- rounding the inclusion of pregnant and lactating women in clinical trials. The Task Force on Research Specific to Pregnant Women and Lactating Women (PRGLAC) wrote a report to the secretary of health and human services and Congress in September 2018, outlining strategies for identifying and addressing gaps in knowledge and research regarding drug use of pregnant and lactating women. Because of this gap in scientific knowledge on the effects of treatments for the health needs of pregnant and lactating women, fair inclusion implies that a boost of research in this population is warranted . While efforts in trial design are discussed, application of AI in this domain remains lacking. The state of maternal healthcare in the USA is currently at a cusp, maternal mortality is increasing despite decreases observed worldwide. In addition, the pharmacological effect of drugs taking during pregnancy still remains largely unknown and underexplored. Clearly better science is needed
and this must go beyond the inclusion or exclusion of pregnant women in clinical trials as suggested by Eke et al. Rather better AI methods are needed that appropriately harness both the existing data in terms of Electronic Health Records (EHRs) and also toxicological and chemical data from the pharmacologics themselves.

RECENT USES OF AI TO UNDERSTAND MATERNAL AND FETAL HEALTH OUTCOMES FROM PHARMACOLOGICS
A major challenge for methods that seek to understand the fetal and maternal consequences of pharmacologics taken during pregnancy is that few randomized controlled trial data exist in humans. The focus tends to be on at-risk populations, such as opioid substitution therapy, HIV prevention therapy, and preterm birth. Concerning the general pregnant population, there is a gap of research on medication use during pregnancy. Researchers must utilize data taken during routine clinical care, and therefore studies are often retrospective outcomes studies. In addition, while data are available in animal models there are major gaps in translating this information to the human context. Therefore, data exists, but these data must be repurposed to answer important clinical questions. This is where AI methods can be especially powerful. Research in ART is represented well due to the fact that pharmacological intervention is necessary for several common ART practices. AI methods have been applied to inform and advise physicians, to predict pregnancy success, to provide optimized treatment, and to understand miscarriage risk. Navigating infertility and ART treatment often takes several treatments and can be cost-prohibitive, and there- fore there is significant focus on the prediction of pregnancy success and applications to improve treatment. Kaufmann et al. applied neural networks to predict success for individual couples about to undergo in-vitro fertilization (IVF) treatment. Neural networks were created using 4 variables: maternal age, number of eggs retrieved, number of embryos transferred, and embryo freeze status (i.e., fresh or frozen). The highest predictive success of the 8 neural networks was 60%, which may be due to the fact that the input information was not sufficient—there is likely an absence of important predictor variables from the data set. It has developed a case-based reasoning system that relies on context-based relevance assessment to assist in knowledge visualization, interactive data exploration and discovery in IVF. This CDSS acts as an advisor to the physician and can help inform the treatment to improve success rate; using 39 attributes, the CDSS suggests the amount of hormonal stimulation for treatment and suggested day for triggering the ovulation. Variables considered and divided into categories for the model include the following: maternal age, previous IVF, intra cytoplasmic sperm injection (ICSI) cycles, grade of each embryo, insemination technique, maternal/uterine receptivity, embryo viability, and pregnancy. Maternal receptivity and embryo viability were only partially observed, and therefore the authors used the Expectation– Maximization (EM) algorithm to estimate parameters. The model predicted the occurrence of pregnancy with an area under the curve (AUC) of 0.72. However, the model requires validation from a prospective study and is possibly a simple model for the complexity of implantation, lacking inclusion of more relevant features. Hassan et al. propose a method to predict IVF pregnancy using a hill climbing feature selection algorithm coupled with automated classification using a variety of supervised machine learning classifiers. Important attributes of the 25 features were chosen by a majority of the classifiers, including maternal age, indication of infertility factor, antral follicle counts, and number of mature eggs. The dataset contained 64 features, 52 of which were female features and the other 12 related to the male. While this method included features of baseline luteinizing hormone, baseline follicle-stimulating hormone, and base- line estradiol, the type of treatment is not used as a feature to estimate success. Their methods are more tailored to screening potential IVF clients to determine success from clinical features. it proposed a web-based system to assist clinicians to provide personalized treatment for sub fertile couples and improve ART outcome , The system relies upon an ANN model and a database system that combines several databases across the health information system, including an IVF database. An example of a set of parameters for the ANN model includes cycle characteristics 24 parameters, couple’s evaluation 12 parameters, female evaluation 30 parameters, and male evaluation 12 parameters. Notably, albumin, gona- dotrophin, and metformin use are listed as parameters, along with cortisolone co-use and stimulation protocol. The proposed system would assist at several points of care during IVF treatment and inform the model to enhance its performance with each record. SVM classifier with a linear kernel was used to predict the recurrent miscarriage and healthy pregnancy classes. An implication of this research is that accurately assessing the risk of recurrent miscarriage associated with a given pair of gametes could improve gamete donor selection and therefore increase pregnancy success rates.
A reoccurring theme from these ART studies is that predicting pregnancy success is complex, and models lack sufficient features for the most accurate prediction. While some models consider pharmacologic interventions as a feature for predicting pregnancy viability, it is often not determined to be an important attribute. A number of ovulation induction treatments are commonly used in infertility treatments: estrogen antagonists, insulin sensitizing agents, gonadotrophins, and GnRH analogs. Further ART research applying machine learning methods could include the type of ovulation induction treatment as a feature.
As a result of the interviewing process, the authors gathered a set of rules. Then, experts reviewed the cumulative rule set and determined whether each rule is a confidence rule or an important rule; while the expert may have confidence that the rule is valid, it may be of little importance. There was significant disagreement between experts and the rule classification required further clarification to the experts, showing that classification boundaries were unclear. Overall, the authors determined six variables that con- tribute to an expert’s decision as to whether or not a compound or agent is a developmental toxicologic hazard: human studies results, animal studies results, whether an active compound is present in the human, physical structure similarity to a known human develop- mental toxicant, mechanism of action similarity to a known toxicant, and whether the compound is a known mutagen or direct cytotoxic agent. There is a need to elicit a set of rules that cover pharmacologic principles, including aspects such as dose amount, absorption, route of exposure, mechanism of action, timing of exposure, and drug/disease interactions. Prevalence of maternal chronic disease has been increasing in the United States. The number of women presenting at hospitalized delivery with 1 or more chronic conditions rose from 66.9 to 91.8 per 1000 delivery hospitalizations between 2005–2006 and 2013–2014. Chronic hyper- tension, chronic respiratory disease, substance-use disor- ders, and pre-existing diabetes are disorders with the greatest increase of prevalence over time. One paper was found relating to chronic disease and pregnancy out- come. Systemic lupus erythematosus (SLE) is a chronic autoimmune disease with unknown etiology, and different clinical manifestations, laboratory signs and prognosis. Pregnancy among SLE-affected women is highly associ- ated with poor obstetric outcomes, namely fetal loss from spontaneous abortion or intrauterine death. Paydar et al. developed a CDSS to predict pregnancy outcomes among SLE-affected pregnant women, namely spontaneous abortion or live birth. Two ANNs were trained based on features selected by a binary logistic regression (LR) model: a multi-layer perception (MLP) model and radial basis function (RBF) model.

We Claims:

1. An implication of this research is that accurately assessing the risk of recurrent miscarriage associated with a given pair of gametes could improve gamete donor selection and therefore increase pregnancy success rates.
2. AI methods have been applied to inform and advise physicians, to predict pregnancy success, to provide optimized treatment, and to understand miscarriage risk.
3. Concerning the general pregnant population, there is a gap of research on medication use during pregnancy.
4. The system relies upon an ANN model and a database system that combines several databases across the health information system, including an IVF data Concerning the general pregnant population, there is a gap of research on medication use during pregnancy.
5. A number of ovulation induction treatments are commonly used in infertility treatments: estrogen antagonists, insulin sensitizing agents, gonadotrophins, and GnRH analogs.
6. While this method included features of baseline luteinizing hormone, baseline follicle stimulating hormone, and base- line estradiol, the type of treatment is not used as a feature to estimate success.
7. The majority of these medications are prescribed and taken without any randomized controlled trials that include pregnant women and this limits the ability for clinicians to understand the potential adverse health consequences both for the mother and the fetus.