IBM this week presented research examining how AI and machine learning can be used to improve maternal health in developing countries and predict the occurrence and progression of type 1 diabetes. In a study funded by the Bill and Melinda Gates Foundation, IBM researchers created models to analyze demographic records from African countries and found “data-driven” associations between the number of years between pregnancies and the size of a woman’s social network with birth results. In a separate paper, another team at IBM analyzed data from three decades and four countries to try to anticipate the occurrence of type 1 diabetes between 3 and 12 months before the typical diagnosis, and then predict its progression. They claim that one of the models accurately predicted progress 84% of the time.
Improving Newborn Outcome
Despite a global decline in child mortality rates, many countries are not on track to meet proposed targets to end preventable deaths in newborns and children under 5 years of age. Unsurprisingly, progress towards these goals remains inconsistent, which is reflected in disparity in access to health services and unequal resource allocation.
With a view to possible solutions, IBM researchers tried to identify features related to newborn mortality “that were recorded in nationally representative cross-sectional data”. They analyzed corpora from two recent demographic and health surveys (from 2014 and 2018) in 10 different sub-Saharan countries and created a model for each survey to classify (1) the mothers who had one in the 5 years prior to the survey Reported birth. (2) those who reported having lost one or more children under 28 days; and (3) those who did not report losing a child. Researchers then examined each model by visualizing the features in the data that influenced the model’s conclusions, as well as how changes in the values of the features might have affected neonatal mortality.
The researchers concluded that in most countries (e.g., Nigeria, Senegal, Tanzania, Zambia, South Africa, Kenya, Ghana, Ethiopia, the Democratic Republic of the Congo, and Burkina Faso), newborn deaths account for most of the loss of newborns makes up children under 5 years of age and that the percentage of newborn deaths has remained historically high despite a decline in deaths under 5 years of age. They found that the number of births in the past 5 years correlated positively with newborn mortality, while household size correlated negatively with newborn mortality. In addition, they said they found that mothers living in smaller households were at higher risk of newborn mortality than mothers living in larger households, with factors such as the age and sex of the head of household appearing to influence the relationship between household size and newborn mortality.
The study co-authors note the limitations of their work, such as the fact that the surveys that are self-reported may omit important information such as health care accessibility and health care search behavior. They also acknowledge that the models may identify and exploit unwanted patterns to make their predictions. Still, they claim to have made an important contribution to the research community to demonstrate that the ensemble’s machine learning can potentially infer newborn outcomes from health surveys alone.
“Our work demonstrates the practical application of machine learning to generate insights through the inspection of black box models and the applicability of using machine learning techniques to generate novel insights and alternative hypotheses about phenomena captured in health data at the population level “the researchers wrote in a paper describing their efforts. “The positive correlation between the number of births reported and neonatal mortality, which is reflected in our results, confirms the previously known observation of the birth distance as a key factor in neonatal mortality.”
Type 1 Diabetes Prediction
A separate IBM team studied the extent to which AI could be useful in diagnosing and treating type 1 diabetes, which affects around 1 in 100 adults during their lifetime. Based on research showing that clinical type 1 diabetes is generally preceded by a condition known as islet autoimmunity, in which the body consistently produces antibodies called islet autoantibodies, the team developed an algorithm that aggregates patients and counts the clusters to be discovered and their profiles determine similarities between different geographical groups.
The algorithm took into account profiles based on types of autoantibodies, the age at which autoantibodies were developed, and imbalances in autoantibody positivity. After pooling the autoantibody-positive subjects, the researchers applied the model to data from 1,507 patients in studies conducted in the United States, Sweden, and Finland. The accuracy of cluster transfer was reportedly high, with a mean of the above 84%, suggesting that the AAb profile can be used to predict the progression of type 1 diabetes regardless of population.
In a related study, the same team of researchers created a type 1 diabetes ontology that captures the patterns of certain biomarkers and uses them together with a model to identify features. The co-authors claim that when applied to the same datasets as the clustering algorithm, the ontology improved predictive performance up to 12 months in advance, allowing predictions to be made of which patients may develop type 1 diabetes a year before they were detected.
It is of course important to note that imbalances in the data sets may have skewed the predictions. A team of British scientists found that almost all eye disease records come from patients in North America, Europe and China, which means algorithms for diagnosing eye diseases among racial groups from under-represented countries are less certain. In another study, Stanford University researchers claimed that most of the US data for studies on the medical use of AI came from California, New York, and Massachusetts.
The co-authors of an audit last month recommended that practitioners conduct “rigorous” fairness analyzes before deploying in order to find a solution to bias. We hope that if IBM researchers decide to use their models at some point, they will take their advice.