Data-Driven Sequential Uptake Pattern Discovery for Family Planning Studies

Abstract

Family planning is a crucial component of sustainable global development and is essential for achieving universal health coverage. Specifically, contraceptive use improves the health of women and children in several ways, including reducing maternal mortality risks, increasing child survival rates through birth spacing, and improving the nutritional status of both mother and children. This paper presents a data-driven approach to study the dynamics of contraceptive use and discontinuation in Sub-Saharan African (SSA) countries. We aim to provide policymakers with discriminating contraceptive use patterns under different discontinuation reasons, contraceptive uptake distributions, and transition information across contraceptive types. We used Demographic Health Survey (DHS) Calendar data from five SSA countries. One recurrent pattern found was that continuous usage of injectables resulted in discontinuation due to health concerns in four out of five countries studied. This type of temporal analysis can aid intervention development to support sustainable development goals in Family Planning.

Introduction

The World Health Organization (WHO) defines Family Planning (FP) as a means to allow individuals and couples to anticipate and attain their desired number of children and the spacing and timing of their births. The effective use of contraceptives can significantly improve the health of women and children, and it has been shown to reduce maternal mortality risks, and improve child survival through birth spacing, as well as improve the nutritional status of both mother and child¹. Contraceptive discontinuation (CD) occurs when the use of a contraceptive is stopped for any reason while the woman is still at risk of an unintended pregnancy. It is estimated that between 20-40% of users discontinue a contraceptive method for reasons other than a desire to become pregnant or no longer needing a method (e.g., post-menopausal)². While a portion of women would switch to a different contraceptive method after discontinuation, many do not, leaving them at risk of an unintended pregnancy. Thus, CD may signal a dissatisfaction with the contraceptive used, and hence it remains an important topic for domain-experts and policy makers in order to improve the efficiency of FP^3–7. While there are multiple studies on FP and contraception, focusing on measuring fertility rates⁸, childbirth spacing⁹, modelling prevalence of modern methods¹⁰, and grouping of calendar episodes by long and short-term contraceptive methods¹¹, not many studies were conducted to identify temporal patterns that lead to discontinuation over time. Our work aims to uncover temporal patterns of contraceptive discontinuation and reasons for discontinuation across countries and subpopulations of women. Such patterns can lead to further research questions regarding the demand and supply chain issues of contraceptives or how to develop robust educational interventions to provide information regarding contraceptive methods and their side-effects.

This work was done in partnership with the Bill & Melinda Gates Foundation (BMGF), specifically with their Family Planning team who advised on research questions of interest, data sources, and provided feedback on proposed methods and evaluation of results. In this paper, we explore two questions regarding contraception use and discontinuation episodes¹: (Q1) Are there any recurrent sequences of contraceptive use and discontinuation across countries? (Q2) What do women transition to when they discontinue or switch between contraceptive methods? To address these questions, a series of events are analyzed in a time window of 12 months using contraceptive calendar data from the Demographic and Health Surveys Program (DHS)¹². Specifically, in each month, DHS provides 1) the type of contraceptive method being used, if any, 2) the events that occurred in that month such as pregnancy, birth, termination, and 3) the reason for discontinuation (if it occurs). While the DHS data provided longitudinal calendar data (for five years) on women’s contraceptive use, the latest 12 months were used to reduce the impact of recall bias. Pattern mining was performed to check for common patterns of events among women that discontinue their contraceptive uptake. In this paper, two major scientific contributions are provided in line with addressing the above research questions. First, discriminatory subsequences are detected and characterized, along with the extent of their discriminatory power, for different types of discontinuation reasons (Q1). Second, we provide an analytic tool to visualize uptake distributions and transition between contraceptive methods in various subpopulations (Q2). The tool enables domain experts to easily visualize contraceptive switching patterns and interact with the obtained results. For the first time, this allows us to automatically compare subgroups of interest using patterns that are unique to a subgroup. Such patterns are valuable to a practitioner to gain a notion into contraceptive uptake time-based patterns and discontinuation trends. The analysis shows that patterns of continuous usage of injectables were a precursor to contraceptive discontinuation due to health concerns in four out of five Sub-Sahara African (SSA) countries studied. Additionally, we observed that such sequences for DHS Kenya (2014), DHS Ghana (2014), DHS Nigeria (2013) and DHS Burkina Faso (2010) are discriminant i.e. statistically unique to women who used injectables.

To the best of our knowledge, this is the first presentation of data-driven and automatic subsequence mining used to understand patterns of contraceptive use in various subpopulations. Furthermore, the proposed dashboard-based interactive and visual analytic tool enables policymakers to explore longitudinal contraceptive dynamics and discontinuation reason via its interface that allows interactive exploration of contraceptive use, distribution, and pattern uptake trends across different countries.

Related Work

Previous studies in this field of research mainly used traditional models, such as decision trees and logistic regressions, to perform stratified analysis to understand the correlations between contraceptive uptake and pre-determined covariates^3–7,13. The majority of these studies are focused on India (the second most populous Asian country) 7 and Nigeria (the most populous African country)^3–5. A few studies were also able to understand the determinants of contraceptive use across multiple countries, particularly Sub-Saharan African (SSA) countries^5,14 and Cahill et al.⁶ analysed the prevalence rate across multiple countries in the world. A summary table is provided in Table 1.

Table 1:

Summary of related works on contraceptive use. Countries: Kenya (KE), Malawi (MA), Madagascar (MD), Rwanda (RW), Cameroon (CA), Chad (CH), Zambia (ZA), Zimbabwe (ZI), Ghana (GH) and Mali (MI). LARC: long-acting reversible contraceptive; LR: Logistic Regression. Data source: Demographic and Health Surveys (DHS).

Authors	Country	Data	Research goal	Method
Azukie et al.3 (2017)	Nigeria	DHS	Identifying predictors of discontinuation	Binomial LR
Michael4 (2017)	Nigeria	DHS	Influence of socio-cultural and economic factors on contraceptive use and desire for less children	Multivariate LR
Larsson & Stanfors13 (2014)	GH, KE, MA, ZA	DHS	The influence of women education and empowerment on contraceptive use	Bivariate and Multivariate LRs
Adedini et al.14 (2019)	MA, Rw, CA, CH, ZA, ZI, GH, MI	DHS	Identifying patterns of LARC	Multinomial LR
Ours	KE, NG, BF, GH, ET	DHS	Mining sequential episodes from calendar data to understand contraceptive uptake patterns and discontinuation reasons in different subpopulations.	Discriminatory Subsequence Mining and Interactive

Most existing works on the analysis of FP programs and identifying key determinants of contraceptive use tend to employ simple modeling techniques, such as regressions. These studies largely ignored the time-based calendar data, which we believe may reveal novel patterns on contraceptive use in different groups. In contrast to these prior studies, we employ sequence mining methods to provide data-driven patterns into contraceptive discontinuation, using the sequential calendar data to mine uptake patterns.

Materials and Methods

In this section, the dataset used for the analyses and the extension of PrefixSpan to analyze contraceptive discontinuation is described. An overview of the proposed approach is shown in Figure 1.

Figure 1:

Overview of the proposed framework that takes temporal data, such as contraceptive calendar data available in DHS datasets, for a given country C_n. There are two main modules that feed the analytics of the dashboard in the framework. First, the Discriminatory pattern mining module is in charge of extracting the top k discriminatory subsequences between two classes defined by discontinuation reason R_m to display on the dashboard. Second, the transition extractor handles preprocessing episode data and probability distribution uptake estimation. This module returns a pair of a probability (Probⁿ_i), transition (Transitionsnⁿ_i) for each contraceptive method CM_k for C_n.

The dataset

The DHS program² has collected, analyzed and disseminated representative data on population, health, HIV and nutrition through more than 300 surveys in over 90 different countries. Nationally representative surveys are designed to collect data on monitoring and impact evaluation indicators important for individual countries and for cross-country comparisons. The DHS calendar¹² dataset contains monthly history of certain key events in the life of the respondent. The calendar collects a complete history of women’s reproduction and contraceptive use for a period of between five and seven years prior to the survey. The exact length of the period covered by the contraceptive calendar varies depending on the duration of data collection, whether the survey overlapped two years and the month in which the respondent was interviewed¹². In the current standard DHS-7 questionnaire for the calendar consists of two values per episode. The first value represent events such as: births, pregnancies, terminations and contraceptive use (twenty six different types of episodes are considered). The second value represents the reason for discontinuation of contraceptive use (nineteen different reasons are considered). For this study, 95, 855 calendars were processed, spanning five SSA countries namely, DHS Kenya (2014), DHS Nigeria (2013), DHS Burkina Faso (2010), DHS Ghana (2014) and DHS Ethiopia (2016) with a total of 1, 150, 260 episodes.

Data-driven Discriminatory Sequence Mining

Sequential pattern mining is a data mining technique that discovers frequent subsequences in a sequence database, such as DHS Contraceptive Calendar data. PrefixSpan¹⁵ is one of such pattern mining techniques and it employs a projection-based, sequential pattern-growth approach for efficient mining of temporal patterns is used. In this approach, a sequence database is recursively projected onto a set of smaller projected databases, and sequential patterns are grown in each projected database by exploring locally frequent fragments only. Several notations for the contraceptive pattern mining problem being investigated are defined below.

Definition 1 Let B = {e₁, e₂, · · · , e_m} be an item base consisting of items, in this case, different types of contraceptive episodes. An itemset is a subset of items if I ⊆ B. A sequence is an ordered list of itemsets, in our case a series of monthly ordered episodes from calendar data. A sequence T = (t₁ → t₂ → · · · → t_n) with ∀ k 1 ≤ k ≤ n : t_k ⊆ B.

Definition 2 Let I ⊆ B be an itemset and T a sequence over B, the cover of I is defined as the set:

$$K_T(I)=k\in 1,\cdot \cdot \cdot ,n|I\subseteq t_k$$ (1)

The cover of an itemset is the index set of sequences that contains all items in I. The value σT = 1/n|K_T (I)| is called relative support of I. The support of I is the fraction of sequences that contain it.

Definition 3 Given the minimum support σ_min ∈ R, 0 < σ_min ≤ 1, the set of frequent itemsets is defined as:

$${\Phi }_T({\sigma }_{\min })=I\subseteq B|{\sigma }_T(I)\ge {\sigma }_{\min }$$ (2)

Definition 4 A sequence s = (s₁, s₂, · · · , s_n) matches a sequence s′ = (s′₁ , s′₂ , · · · , s′_m) if there exists j₁ < j₂ < · · · < j_n such that s_i = s′_ji, the function is defined as match(s, s′).

In this study, the goal of using sequence mining is to identify contraceptive uptake patterns that may be unique to one cohort of women (e.g., women that discontinued due to health concerns). Extracting such sequences do not only shed information regarding uptake patterns to domain experts, but can also have predictive power —i.e., be able to predict that a discontinuation may occur in k steps if a particular pattern is being observed. To identify such unique or “discriminatory” sequences, the PrefixSpan¹⁵ is extended in several ways. First, we need to look for differentiating sequences between two classes, namely women who discontinued due to health concerns and those who did not discontinue. To address this, PrefixSpan was extended to mine for patterns that are different between the two classes of interest. There are two parameters that need to be set, namely σ_min (Def. 3) and w_max. The first parameter expresses a threshold on how often a particular pattern occurs, in terms of the minimum percentage across all the sequences in DHS calendar data for a given country. The patterns identified are sequential in nature and are subsequences of the contraceptive uptake patterns. w_max defines the time window to be analyzed. PrefixSpan reporting method is extended such that each pattern is tagged with various metrics. While these are familiar concepts in association rule mining, their definitions to mine sequences with discriminatory power are slightly modified in this work. The two key metrics are support and lift. Support left and right, is the support σ_T (Def 2) of a pattern T, for left and right datasets, respectively. $Lift(T,left,right)$ is the ratio of = $\frac{{\sigma }_T^{left}}{{\sigma }_T^{right}}$, defining how often a pattern occurs in the left dataset compared to the right. When this number is very large, it indicates that the pattern is unique to the left dataset. With these definitions, the problem of finding discriminatory patterns can be formalised as follows: Let $p_s=\frac{\mathrm{Pr}(match(s,S_1))}{\mathrm{Pr}(match(s,S2))}$ the ratio of finding a given pattern, where S₁ and S₂ are two collections of sequences (e.g., those that continue contraceptive use vs. those that don’t). Let $d_s=\max (p_s,\frac{1}{p_s})$ denote the discriminatory power of sequence s with respect to the two collections of sequences S₁ and S₂, where a larger d_s ≥ s is a more powerful discriminator.

Given the two collections of sequences S₁ and S₂, learn as a collection of sequences S(|S| << |S₁|, |S₂|) such that ∀_s ∈ S, d_s with respect to S₁ and S₂ is greater than a given threshold (th). We note that pruneByDominance is a function to eliminate those patterns q that are obtained by augmenting an existing pattern p, where p is shorter or more general than q, and has a higher confidence of predicting a class than q¹⁶. Algorithm 1 shows the pseudo-code for discriminatory sequence mining (DSM).

Algorithm 1: A pseudo-code for the proposed Discriminatory Sequence Mining.

Experimental Analysis and Results

In this section, we present a set of experiments conducted to answer the two research questions outlined in the Introduction, by analysing contraceptive use and discontinuation from different angles to provide a holistic view of the problem landscape to domain experts. First, motivated by (Q1), the discriminatory sequence mining algorithm is used to mine frequent temporal patterns of contraceptive use across all countries. Second, regarding (Q2), contraceptive transitions is studied to answer the following questions: how long an individual continues to use a contraceptive option? And what they transition to when a method is discontinued?

Experimental setup for Q1: Discriminatory subsequence mining for fine-grained pattern extraction Data Preprocessing

For sequence mining experiments, more than 95, 855 contraceptive calendars corresponding to women population from five Sub-Saharan countries were analysed, i.e., DHS Kenya (2014), DHS Nigeria (2013), DHS Ghana (2014), DHS Ethiopia (2016), and DHS Burkina Faso (2010). Retrospective reporting of contraceptive use and discontinuation rely heavily on the ability of respondents to accurately recall events. To reduce recall bias, the 12 months prior to a discontinuation event are used in this analysis, and the three months immediately before the survey is removed to account for under-reporting of first trimester pregnancies at the time of the survey¹⁷. For all the experiments presented in this section, calendar data from individuals who wanted to become pregnant are excluded (DHS Code 2).

As described in the Methods section, two parameters are needed to train the Discriminatory Sequence Model, namely the time window to be analyzed (w_max) and the minimal support required (σ_min). These parameters were tuned in consultation with domain experts, understanding that smaller minimal support within a time window provides more sequences needing to be examined for interpretation, and a shorter window provides patterns including short-term contraceptive methods. A time window of w_max = 12 months is constructed to use the full calendar year extracted and capture both long- and short-term contraceptive methods. We selected 0.3 as a minimum fraction of the population required for the patterns to be significant for minimal support. In order to work with DSM, two groups for evaluation were determined. For example, Table 2 shows the discrimination patterns of calendar data that contain discontinuations due to health concerns as opposed to the rest of the population. The reason for discontinuation is extracted from the discontinuation column. To evaluate each frequent pattern, Support Left, Support Right, and Lift are calculated. All experiments were performed in a desktop machine (2.9 GHz Quad-Core Intel Core i7, 16 GB 2133 MHz LPDDR3) and the implementation was done with the Scientific Python Stack¹⁸.

Table 2:

Discriminatory subsequences mined in the subpopulation of women who discontinue due to health concerns (DHS Discontinuation Code 4), as opposed to the rest of the population that discontinue for any other reason. Injectables (DHS Code 3), non-use (DHS Code 0)

Country	Subsequence	Supp. Left	Supp. Right	Lift
KE	3 → 0	0.33125	0.00909	36.429
NG	3 → 3 → 3 → 0	0.33644	0.00005	561.28
NG	3 → 3 → 0 → 0	0.34579	0.00005	663.40
GH	3 → 3 → 0 → 0	0.33888	0.10357	3.2715
BF	3 → 3 → 0 → 0 → 0	0.31896	0.14345	2.2233
ET	3 → 3, → 3 → 3 → 0	0.30508	0.27232	1.1203

Results

All subsequences found under health concern discontinuations contained episodes of injectables (DHS Code 3) as the contraceptive method (Table 2). Additionally, patterns for DHS Kenya (2014), DHS Ghana (2014) and DHS Burkina Faso (2010) are discriminant (See Support Left, Support Right and Lift metrics) while for DHS Ethiopia (2016) the pattern containing injectables was not discriminant for the health concern discontinuations. Additional discriminatory subsequence experiments with other reasons for discontinuation can be seen in Table 3. We can observe that in both DHS Kenya (2014) and DHS Ghana (2014) discriminative patterns were found with consecutive use of Rhythm (DHS Code 8) as a contraceptive method in the group of calendars that reported to discontinue due to becoming pregnant while using a contraceptive method (DHS Discontinuation Code 1). For sequence mining visualization, a combination of directed graphs and bar plots (for support metrics) are used, with an example provided in Figure 3.

Table 3:

Examples of discriminatory subsequences with particular discontinuation reasons. DHS Code Discontinuation Reasons (DR): Became pregnant while using a contraceptive method (DHS Discontinuation Code 1), wanted to become pregnant (DHS Discontinuation Code 2). Episodes: Rhythm (DHS Code 8), Pregnancy (DHS Code P).

Discriminatory reason	Country	Subsequence	Support Left	Support Right	Lift
(1) vs (2)	KE	8 → P → P	0.26785	0.07596	3.5260
(1) vs (2)	KE	8 → 8 → P	0.27142	0.07658	3.5440
(1) vs all other DR	GH	8 → 8	0.30693	0.07666	4.0034

Figure 3:

Discriminatory subsequences with support for Side effects/Health concerns (4) versus other discontinuation reasons (X) for DHS Ghana (2014). (A, B): The support value of the pattern for the left dataset and the right dataset respectively, are aggregated for all individual subsequences. i.e., the mean and standard deviation are computed. (C, D): Each episode e.g. Injectable, is repeated n number of times where n is indicated using the corresponding color intensity on the color bar (E). A subsequence may consist of a single repeated episode as in (D) or multiple episodes linked with an arrow as in (C).

Experimental setup for Q2: Contraceptive Transitions

Probability distributions of contraceptive transitions

To provide an overview of contraceptive use across each country based on the calendar data, the probability distributions per contraceptive type is estimated. The domain expert can define the time interval to use for the estimations, which can be a specific year of the calendar data or the totality of years available. To determine how long an individual used a particular contraceptive before switching to other methods or discontinuing to non-use, the frequencies of consecutive months of use for each specific contraceptive per country was calculated. Given I the itemset of type of episodes, D_i the set of sequences of calendar data for the country i, we extracted subsequences with one type of episode j ∈ I − {B, P, T, 0}. The length for each subsequence was then calculated, excluding episodes that do not correspond to a contraceptive method. As a result, a set of vectors is constructed V_ji = (|T₁|, · · · , |T_n|), where j ∈ I and i ∈ D |T_i| is the length of each subsequence found for that contraceptive over all calendar data in D_i. The Kernel Density Estimation (KDE) algorithm¹⁹ is used to estimate the probability density function applied to {V_ji|j ∈ I − {B, P, T, 0}}. An example of interactive violin plots for consecutive months of contraceptive use distributions for DHS Kenya (2014) with six different contraceptive methods (five modern methods and one traditional method) is shown in ^{Figure 2}. This type of visualisation provides the domain expert with the tools to compare distributions across countries and methods. The probability of an individual using a particular method of contraception for a number of months is also estimated. Domain experts may find such analyses helpful for visualizing the switching patterns between different contraceptive; not simply who discontinued or why, but also some insight on what the type of transition was (i.e., what type of contraception the person used next, if any). Furthermore, the probability distribution of consecutive episodes may allow experts to see and compare length of use in ways that show patterns not apparent from the usual “average months of use” that is commonly used.

Figure 2:

Interactive Violin plots for different contraceptive methods across subsequent months in the calendar data from DHS Kenya (2014). The x-axis represents the contraceptive method, and the y-axis the number of consecutive episodes that a particular method was used.

After studying the trends for consecutive months of use for each country, a simple one-to-one pattern transition is provided. Using the subsequence already extracted in the previous section, ordered pairs such as a set of tuples P_k = ((T₁, T₂), · · · , (T_n−1, T_n)) are formulated, where T_i is a subsequence of one type of contraceptive j with j ∈ I, k ∈ D and (T₁, T₂) is a transition pair from each subsequence found to other contraceptive methods in D_k. Examples of these transitions can be seen in Figure 4. In each Sankey plot, transitions from one type of episode, in this case an Injectable and Rhythm episode, to all the other types of episodes, may be observed. For women that transit, domain experts would find it useful to know what they transition to, in order to better understand contraceptive uptake. These one-step transitions are visualized using a Sankey plot, wherein the contraceptive of interest is selected as the ’source’ of the flow and plot the amount of transition (indicated by the flow size) to all other ’destinations’ i.e., contraceptive methods. The Sankey plots in Figure 4 illustrate the transitions from injectables in DHS Burkina Faso (2010). This analysis was repeated for all five countries and the results are summarized here for brevity.

Figure 4:

Examples of interactive Sankey plots for transitions across episodes, in this case from Injectable and Rhythm episodes, to all the other types of episodes. We can observe from the visualization that Injectable method of contraception has high rates of transition to non-use. As for Rhythm episode we can observe high rates of transition to pregnancy and non-use.

Results

The percentages of transitions to non-use (no contraceptive used) constitute the largest proportion among all the countries studied: 85.36% in DHS Burkina Faso (2010), 57.39% in DHS Kenya (2014), 76.7% in DHS Ethiopia (2016), 70.42% in DHS Nigeria (2013) and 75.72% in DHS Ghana (2014). Importantly, injectables as a method of contraception had high rates of transition to episodes of non-use in all countries. In contrast, episodes of Rhythm use showed high rates of transition to both pregnancy and non-use in all countries. Figure 4 shows the transitions from injectables and Rhythm in DHS Burkina Faso (2010). The percentages to pregnancy episodes were: 57.38% for DHS Burkina Faso (2010), 59.67% for DHS Kenya (2014), 34.38% for DHS Ethiopia (2016), 55.36% for DHS Nigeria (2013) and 67.18% for DHS Ghana (2014). And the percentages to non-use were 34.43% for DHS Burkina Faso (2010), 16.57% for DHS Kenya (2014), 53.12% for DHS Ethiopia (2016), 30.36% for DHS Nigeria (2013) and 24.43% for DHS Ghana (2014). An interesting result that emerged was the bimodal distribution of Rhythm users compared to other methods, a finding that was not evident to the team until the episodes were visualized in this way (Figure 2). This finding implies that while some women briefly try the Rhythm method, a significant portion of them use this method for extended periods of time. Given that Rhythm has a high transition rate to unintended pregnancy, this insight draws attention to domain experts who aim to reduce the risk of unintended pregnancy.

Discussion

Our goal with this work was to illustrate the potential of sequence mining techniques to provide a data-driven approach to understanding contraceptive dynamics using the DHS survey data. This may inform policymakers and health workers as they address unmet needs in contraceptive usage. As such, this work provides domain experts with the ability to visualize and compare frequency distributions and discontinuation flows across different methods, sub-populations, and countries. Our analysis shows that patterns of continuous injectables episodes were a precursor to contraceptive discontinuation due to health concerns in four out of five datasets studied. These results need further examination to understand why these patterns appear and how the subpopulation that exhibits these sequences of episodes can be better characterized. Nonetheless, from the research literature, we know that side effects of modern family planning methods, either experienced or anticipated, have been identified as a common reason women choose not to start or discontinue contraceptives. Fear of side effects may occur when a woman or someone she knows has experienced side effects with a method when rumors are considered as facts, or rare complications are exaggerated^20–23. The temporal patterns found in this work complement existing research while providing a time-window that potential interventions can be made. For example, targeted education plans for subpopulations characterized by a pattern of contraceptive use identified result in discontinuation.

In another experimental setup, the domain experts aimed to create a “super dataset” by combining calendar data from multiple countries so as to perform a stratified DSM analysis such that there is sufficient sample size in rare classes. However, discriminatory patterns were hardly extracted from the combined datasets from different countries, which reflects the need to avoid blind merging of different datasets, with different years, countries and/or cultures. An approach of this kind will benefit from taking into account sub-populations that show some similarity in terms of their FP uptake or other factors. This has opened up a new research question: can we identify sub-populations that are “similar” enough to be able to combine their data, thus enabling sufficient data sizes for stratified analysis?

Conclusions

Our proposed approach provides subject-matter experts with insights based on data-driven automatic sequence mining and interactive visualization tools. This work analyzes a series of episodes in a contraceptive uptake pattern and identifying which sequences lead to discontinuation, to what extent these sequences are discriminatory, and whether these sequences are common across sub-populations. The proposed approach was tested across large datasets such as DHS Survey and Contraceptive Calendar Data. As more reliable calendar information becomes available in the future, such as PMA202024, this work demonstrates the potential to provide information to policymakers regarding contraceptive use and discontinuation for a given subpopulation. After finding contraceptive patterns on reliable data, subpopulations with the same patterns may be identified to perform targeted interventions. These interventions can range from educational material regarding a particular method to guarantee access to contraceptive methods after a given time-window.