Actionable Recourse via GANs for Mobile Health

Maternal and neonatal mortality remains a pressing problem in global health. Each year, nearly 300,000 women and 5 million newborns die of causes directly related to pregnancy and childbirth (Fund, 2020; UNFPA, 2020). In addition, nearly all newborn deaths occur in low- and middle-income countries, 80% of which are preventable and treatable by cost-efficient interventions (WHO and UNICEF, 2014; Lawn et al., 2016).

Digital tools such as the App can help build the capacities of skilled birth attendants(Lund et al., 2016; Olusola Oladeji and Oladeji, 2022). The App is divided into clinical modules covering key topics related to maternal and newborn health, including both normal labour and birth as well as guiding healthcare workers through common obstetric and newborn complications. Each module consists of educational videos, easily referenceable action cards, drug lists, practical procedures, and a series of tests that increase in difficulty to assess the depth of knowledge and skills acquired by users.

As defined, recourse must have the following four characteristics: First, recourse actions must be actionable, i.e., they must use features that users can change (Joshi et al., 2019; Karimi et al., 2020; Ustun et al., 2019). For example, recourse may be provided to a user for increasing their lifetime in the app, but not by changing the date of first use. Second, recourse actions must be realistic (Wachter et al., 2017), following real-world directional restrictions of features (e.g., preservation of the monotonically increasing nature of accumulated metrics). Third, recourse actions must be feasible (Ustun et al., 2019) and not prohibitively expensive to execute (e.g., increasing the number of modules completed weekly by a reasonable amount). Finally, recourse actions must obtain the desired predicted outcome, meaning that the implementation of such actions must result in the desired predicted outcome.

Survival analysis models have historically been used in medical and biological research for estimating life expectancy (Clark T.G and D.G, 2003; Fleming and Lin, 2000; Quinlan, 1986). However, the survival analysis toolbox provides methods generalizable to modeling and predicting the time to an event of interest in the presence of right censoring. CSFs (Wright et al., 2017) are a variant of the random survival forests (RSFs) (Ishwaran et al., 2008) that recursively partition the feature space pertaining to the splitting criteria through linear rank statistics. This method minimizes the bias typically encountered in RSF predictions.

In this context, CSFs can be used to predict how long users will remain active before they churn (i.e., the number of days between the first and last login) (Periáñez et al., 2016; Bertens et al., 2017; Kim et al., 2018; Chen et al., 2019; Olaniyi et al., 2022). We define churn as a number of consecutive days without in-App activity, that is determined following Olaniyi et al. (2022).

We use the data pertaining to a sample of users of the App between January 1, 2018 and October 1, 2021, from countries with the maximum App use (India, Myanmar, Ethiopia…). We use a train/test split of 50/50, resulting in datasets of 730 and 729 users, respectively. The in-App logs are processed into daily user metrics, such as the number of daily sessions, time spent using the App, days since last login, and similar values for specific e-learning contents to characterize the user activity and behavior. The metrics are used to construct features by applying statistical operations (mean, max, normalization, etc.) over different periods, resulting in 181 features.

First, the CSF models are trained and used to predict days to churn for the users. Next, the predicted days to churn are converted to a binary label predicting whether the user will have a lifetime of at least 90 days.

Next, we input feature data from users that churned in less than 90 days to $G$ and train it iteratively via gradient descent. We add the generator output and original input to construct the counterfactuals of users that achieve a predicted lifetime of at least 90 days and appear realistic to the discriminator. The discriminator is trained to distinguish users that will churn in 90 days or more and those that have implemented recourse actions to extend their predicted lifetime to at least 90 days.

Each CounteRGAN is trained for up to 600 iterations but saved at checkpoints at which sufficient discriminator performance is observed. These checkpoints correspond to an accuracy less than or equal to 0.55 for the pre- and post-recourse users. Low accuracy on pre- and post-recourse data indicate that the post-recourse data are difficult to distinguish and thus more realistic.

The CSF and GAN performances are evaluated using the following metrics:

• •

  Accuracy: $\frac{1}{n}\sum_{i}^{n}\mathds{1}(y_{i},\hat{y_{i}})$.

  • –

    Model Accuracy: $n$ is the total number of users in a dataset (training or testing, $y\in\{0,1\}$ or $y\in\{0\}$, respectively). The accuracy indicates the model performance in terms of the number of predictions that match the true observed label.

  • –

    Discriminator Accuracy: This metric indicates the discriminator performance in terms of the number of predictions that match the true original data or data augmented by the generator.

  • –

    Classifier Accuracy: $n$ is the total number of $y=0$ users that are augmented sufficiently by the generator to obtain a $\hat{y}=1$ prediction.

• •

  Percent Denied: $\frac{1}{n}\sum_{i}\mathds{1}(\hat{y}(x_{i}),0)$, where $n$ is the number of applicants in the dataset (training/testing). The metric indicates the proportion of applicants predicted to churn in less than 90 days, thereby qualifying for recourse.

• •

  Percent Successful Recourse:
  $\frac{1}{||\hat{y}(x_{i})=0||}\sum_{i\in\hat{y}(x_{i})=0}\mathds{1}(\hat{y}(x_{i}+a_{i}),1)$. This metric indicates the proportion of applicants predicted to churn more than 90 days after adopting the recourse actions recommended by the generator.

• •

  Mean Cost of Successful Recourse Actions: $\frac{1}{||\hat{y}(x_{i}+a_{i})=1||}\sum_{i\in(\hat{y}(x_{i}+a_{i})=1)}||a_{i}||^{2}$. This metric indicates the average $L_{2}$ distance for the feature change for applicants who successfully achieve the desired predicted outcome.

• •

  Cumulative Cost of Denied Recourse: $\sum_{i\in(\hat{y}(x_{i}+a_{i})=0)}||a_{i}||^{2}$. This metric indicates the total $L_{2}$ distance for the feature change for applicants who do not achieve their predicted desired outcome.

• •

  Average Clock Time: $\frac{1}{n}\sum_{i}T_{end}(x_{i})-T_{start}(x_{i})$, where $n$ is the number of users in the dataset (training/testing), and $T$ is the clock time at which the recourse calculation starts and ends. This metric indicates the average time (in seconds) required to calculate a user’s recourse actions to estimate the corresponding counterfactual.

The initial classifier performance over all users and those who churned in less than 90 days is presented in Table 2. As expected, the model performance is enhanced with the forest size, from 0.7177 for all users for model_1 to 0.8709 for model_20. The percent of denied users is consistent across models (Table 3), increasing slightly from 44.96% for model_1 to 52.34% for model_20. The lack of significant variation is attributable to the binarization of the survival prediction.

Low discriminator performance on both real and fake data signifies realistic “generated” (transformed) data (Table 2). In addition, for post-recourse applicants, the model performance is enhanced with the forest size. This phenomenon likely occurs because higher accuracy CSFs can provide more precise feedback to the GAN, allowing it to better learn the data manifold and differentiate classes. Therefore, model_20 GANs represent the optimal platform.

We analyze the time required by different models to compute the recourse by comparing the mean clock time (in seconds), as indicated in Table 3. On the whole, recourse generation via GANs is faster than that based on RGD. This faster runtime is particularly salient in low- and middle-income countries, in which the internet connectivity may often be unreliable or user computational resources may be severely limited.

Table 3 summarizes the values of the percentage of users provided $C(x_{i}+a_{i})=1$ and denied recourse $C(x_{i}+a_{i})=0$ after implementing a CounteRGAN-generated action. Although the percentage of recourse-denied users is consistent across models, the percentage afforded recourse increases for larger forests. This result supports that CounteRGANs trained on more accurate models may be able to provide more effective recourse. In addition, both models_5 and _20 provide more effective recourse than that of counterfactuals obtained via RGD.

We examine the cost of recourse as another auditing mechanism. We estimate the cost of recourse with $L_{2}$ distance across effective and ineffective recourse. Table 3 presents the mean cost of an effective recourse action and cumulative denied recourse cost across models. Model_1 has the lowest mean cost of successful recourse actions, followed by models_20 and _5. However, the cumulative denied recourse cost does not exhibit the same trend: it increases from model_1 to models_5 and _20, thereby highlighting the differences in the cost of recourse across user groups. Figure 3 shows the differences in the cost distributions of effective (left) and ineffective (right) recourse. The cost of effective recourse is similar across models. However, the cost of ineffective recourse differs across models, with model_20 associated with a wider range but lower frequency.

This finding demonstrates that recourse may be providing disproportionate gains and losses across those who received effective and ineffective recourse. In addition, the comparison of recourse via CounteRGANs and RGDs shows that RGDs have a lower cost, in terms of both the mean cost of effective recourse action and cumulative denied cost. The effects of lower efficacy recourse must be considered as it may have negative downstream effects such as lack of accountability, lack of trust, and potential disengagement with the App.

To demonstrate the utility of recourse as an informational tool, stakeholders can examine the distributions of the cost of recourse across true and predicted outcomes. Figure 4 shows two histograms of the cost of recourse associated with a particular feature: the maximum number of actions taken in the last 15 days. The plot to the left shows the distribution across effective and ineffective recourse actions. The plot to the right shows the corresponding distribution across true outcomes.

The cost distributions differ when separated by the predicted outcome and true label. This finding indicates that this feature may not be causally related to the true outcome and may thus have an inflated estimation effect on the true outcome. Examining these distributional differences across important features can help audit the model trust, especially as it is used to provide recourse suggestions to users.

All data used in this analysis are derived from the Safe Delivery App logs and belong to the Maternity Foundation. For inquiries regarding the use of these data, please contact the Maternity Foundation at [email protected]. The code used is available at https://github.com/benshi-ai/paper_2022_ml4h_recourse_gans.

All the analysis were performed with Python version 3.7.13 on Google Colab, using the following packages from PyPI: pysurvival 0.1.2 (Fotso et al., 2019–), alibi 0.7.0 (Van Looveren et al., 2022), and tensorflow 2.8.2 (Abadi et al., 2015).