 Family planning has emerged as a crucial component of sustainable global development and is essential for achieving sustainable universal health coverage. Specifically, contraceptive use improves the health of women and children in several ways, including reducing the maternal mortality risks, increasing child survival rates through birth spacing, and impacting the nutritional status of both mother and children. In this presentation, we're going to talk about data-driven insights on the dynamics and determinants of contraceptive use and discontinuation. This is work done in collaboration with the Bill and Melinda Gates Foundation. Unwork addresses the problem of family planning, which is the crucial sustainable development goal. Specifically, we apply artificial intelligence techniques to better understand the issue of contraceptive discontinuation. In 2020, 40% of women discontinue a contraceptive method due to some dissatisfaction, which in turn can lead to high rates of pregnancy. To address this issue, we build an analytics toolkit for policy makers and healthcare practitioners to gain insights into the dynamics of contraceptive uptake behavior. We do this by analyzing episodes of contraceptive use and discontinuation so practitioners can better understand which subpopulations are most prone to discontinuation. Specifically, we explore three questions. What do women transition to when they discontinue or switch between contraceptive methods? Are there any recurrent sequences of contraceptive use in discontinuation across countries? Can we go beyond covariate analysis to establish a causal effect of particular contraceptive for a specific discontinuation reason? This platform was designed to explore contraceptive use in countries with available DHS survey and calendar data information. The FP platform provides the following facilities. First, automatically characterize subgroups of interest using patterns that are unique to a subgroup, that is, discriminatory, providing the ability to automatically generate hypothesis for various subgroups. Second, estimate the causal effect of a particular contraceptive method on discontinuation across different countries. This technique enables us to systematically verify the hypothesis generated in the first step. Finally, an integrated dashboard with visualization techniques to make these machine learning insights accessible to a domain expert to explore discontinuation and contraceptive use. First, a quick overview of the data. While the DHS data provides us with five years of calendar data for women's contraceptive use, we use the most recent 12 months to reduce the recall bias. We examine a total of 95,000 records. We wanted to understand the typical length of time a woman uses a particular contraceptive method before switching or discontinuing. Depending on the method of contraception used, there is a probability distribution for how long a person will continue with the same method. This gives us information about the adherence to each contraceptive method and that they should not be treated with a unique time window. Next, we provide one-to-one transition visualization using Sankey plots. Here, we see a large proportion of transitions from injectables to non-use. So we need to further study the reason for high rates of discontinuation with injectables. To do this and to delve deeper into this behavior, we treat the monthly episodes of calendar data as events and provide discriminatory subsequence mining, which is an extension we create to the prefix span algorithm to be able to discriminate between two classes of interest and mine frequent subsequences. These subsequences gives us an understanding of longer term contraceptive behavior of women and provide an interpretable way for the subject matter experts to delve deeper into what happened. We compare women in Kenya who wanted to become pregnant versus those who became pregnant while using contraceptives. And what we found is that 25% of the women who became pregnant while using contraceptives exhibit the following sequence of use, namely non-use, followed by the rhythm method, followed by pregnancy. The number of repetitions of each of these steps is shown in the directed graph here and anyone matching that pattern is 3.1 times more likely to get pregnant while using contraceptives in comparison to all other patterns combined. It is interesting to be able to build. We are driven temporal windows when we have more reliable calendar data, for example. But even now, using the DHS calendar data as a proof of concept for these machine learning method. The ability to identify discriminative subsequences that are explainable by a domain expert is immensely beneficial. However, such subsequences alone cannot provide the ability to choose an intervention as there is no insight on the effect of those decisions that will be made by an expert. For example, let's assume we uncover that consecutive months of use of a particular contraceptive method is a precursor to discontinuation due to health reason. We cannot conclude from this if contraceptive use was the cause for discontinuation or if it was a spurious correlation. Also, we cannot determine the extent of the causal effect. For example, what is the increase in likelihood to discontinue due to health concerns from the use of a particular method? These drawbacks motivate us to explore causal analysis of contraceptive discontinuation. Here we set up the causal hypothesis, define the outcomes based on discontinuation reasons and the treatment based on whether injectables was used or not. Then we use inverse probability weighting with stabilized weights. For each country, we calculate the average treatment effect and show these as the percentage causal risk difference in the figure on the left. For all countries, the results suggest that the use of injectables indeed has a causal effect on discontinuation for health reasons, although to varying degrees. For example, we show that a woman in Kenya is five times more likely to discontinue injectables due to health concerns compared to a woman in Ethiopia. Such insights are very actionable by a policymaker. So having looked at our decision platform, let us now take a look at the dashboard which allows an expert to visualize these contributions. Note that this dashboard is continuously updated with preloaded datasets from different countries. To the right, we can see the frequency of consecutive months using a specific contraceptive. We have the classical box plots, but also the probability distributions for how long a person will stay in a particular contraceptive method. We can also select only particular contraceptive methods to analyze more closely. When we move to these continuation flows, we need to select the sorts of episodes that we want to study. For example, injectables. And we want to see all the other types of events that they transition to. As we can see, these are all the episodes that we have for injectables in Kenya and what do they transition to? 57% of the episodes transition to non-use, 13% to pill, 11% to pregnancy, and 9% to implants. If we have any specifics that we want to show, for example, non-use or pregnancy, we can filter them in this label. For the discriminatory subsequence, we have a few preloaded experiments. For example, we were analyzing different discontinuation reasons. So we need to select from this dropdown. For example, let's see has been away. So what we did is build two groups. First all the calendars that have this as a reason for discontinuation has been away and all the rest of the population. So the discriminatory subsequence that appear is that in the group that has a discontinuation reason for has been away, we found the patterns from condom to non-use. If you can see in the right, the color means for how many consecutive episodes we found condom. In this example will be two consecutive months of condom and then transition to non-use. Other examples is more than six consecutive episodes using condom and also condom to non-use. To the left we can see courage left and right. This for example means that 31% of the calendars that we found with a reason for discontinuation of has been away have all these patterns.