benshi.ai

Our mission is to improve the health of individuals and populations in resource-poor countries through advanced AI applications

Research

Healthcare teams may use machine learning recommendations to support clinical decision-making, to triage patients and identify those requiring close follow-up, or to provide portable diagnostics for non-specialist health workers in off-site locations.

benshi.ai focuses on providing personalized, data-driven insights to medical care teams and individual patients. Our advanced AI research does not just lay dormant in servers -- we develop ML products to benefit real-life communities. In addition, we support projects in assessing the effect of decisions and the impact of interventions.

Reinforcement learning

Reinforcement learning formalizes trial-and-error methods of learning and sequential decision-making

In Reinforcement Learning (RL), agents learn how to interact with their environment. They have the ability to learn the best sequence of actions, taking into account the current characteristics of each individual, their previous history, and their stage in the process. RL is particularly well-suited to assessing healthcare decision-making processes.

It can be especially useful in digital contexts, experiment design and assessment (even with inherent time delays), as well as for recommendation systems. RL problems can be solved using model-free learning methods that can be value-based (such as Q-learning or SARSA), policy-based (e.g. policy gradient), or combine elements of both (e.g. actor-critic). Another approach is to use model-based or planning methods, such as dynamic programming.

Reinforcement learning

Causal inference

Causal inference is key in healthcare for deciding which interventions are successful, and assisting in empirical decision-making

Causal inference (CI) compresses a set of methodologies to understand the causal relationship between variables, going beyond most ML models that are based on learnt correlations from data.
When estimating the effect of a new intervention, methods based on only correlations may be biased, because “correlation does not imply causation”. CI tackles this issue.

CI frameworks include nonparametric structural equations, counterfactuals analysis and graphical models. To make ML models generalizable, some of these CI strategies can be incorporated in ML models: causal decision trees, causal reinforcement learning, bayesian regression trees and ensemble learning models for estimating individual treatment effects.

Causal inference

Experimental design

Experimentation is the only way to assess whether our ML-generated predictions and recommendations have produced the intended effect, or whether other interventions have been effective

Experimental design is the collection of statistical techniques used in the design and analysis of experiments, which allow for an empirical estimation of the impact of controlled interventions.

Strictly speaking, experiments require complete randomization, which is unfeasible in many cases due to practical and/or ethical reasons. Nonetheless, we can still estimate the causal impact in these cases, using bayesian experimental design or quasi-experiments. Here too ML can come in handy, in particular to assist in the matching of a non randomized control group. When experimentation is not possible, target trials can be used instead, emulating randomized controlled experiments with observational data.

Experimental design

Dynamic predictive modeling

Predictive modeling is the task of solving classification or regression problems in order to provide personalized predictions and recommendations

We focus on ML methodologies that take into account the sequential dynamic nature of most of the variables of interest.

These include: survival analysis, recurrent and convolutional neural networks, autoregressive time series forecasting, Kalman filtering, deep time series (deepAR, deep Kalman filter, deep state space models, etc), reinforcement learning and sequential recommender systems.

Dynamic predictive modeling