Analytics for Improved Healthcare

The application of artificial intelligence in healthcare requires a team science approach. A diverse set of expertise, perspectives and lived experiences are required to understand the various ways bias lurks in the data – from bias introduced by sampling selection (who made it to the database, who didn’t, and what’s the impact on downstream models), variation in the frequency of measurement that is not explained by the disease or patient phenotype (aka “shortcut” features in medical images), technology that performs differently across patient subgroups (e.g. pulse oximetry, wearable sensors optimized around fit individuals), etc. Data bias is the roadblock to realizing the promise of machine learning. Algorithmic bias is not just about evaluating model performance across patient subgroups post hoc. The goal is to ascertain that the model does not learn from features that should not affect decision making. Offering chemotherapy should not depend on whether a patient is on Medicaid or has a private insurance, predicting job performance should not be informed by the gender of the applicant, optimizing treatment for sepsis should be not be confounded by the use of infrared sensing technology. This is much easier said than done because of the discovery that computers can easily learn sensitive attributes that the human eye does not see. Using real world data to evaluate the models makes this extremely challenging. Excellent model accuracy means existing outcome disparities are fully encoded in the algorithms.

The use of telemedicine has increased rapidly over the last few years. To better manage telemedicine visits and effectively integrate them with in-person visits, we need to better understand patient behaviors under the two modalities of visits. Utilizing data from two large outpatient clinics, we take an empirical approach to study service incompletion for in-person versus telemedicine appointments. We focus on estimating the causal effect of physician availability on service incompletion; physician availability is likely to affect the likelihood a patient leaves without being seen behavior but only for patients who show up for the appointment. That is, physician availability should not impact the likelihood of patient no-show. We introduce a multivariate probit model with instrumental variables to handle estimation challenges due to endogeneity, sample selection bias, and measurement error. Our estimation results show that intra-day delay increases the telemedicine service incompletion rate by 7.40%, but does not have a significant effect on the in-person service incompletion rate. This suggests that telemedicine patients may leave without being seen when delayed, while in-person patients are not sensitive to intra-day delay. We conduct counterfactual experiments to optimize the intra-day sequencing rule when having both telemedicine and in-person patients. Our analysis indicates that not correctly differentiating the types of incompletions due to intra-day delays from no-show behavior can lead to highly suboptimal patient sequencing decisions.

Human donor milk is considered important nutrition for millions of infants that are born preterm each year. Donor milk is collected, processed, and distributed by milk banks. The macronutrient content of donor milk is directly linked to infant brain development and can vary substantially across donations, which is why multiple donations are typically pooled together to create a final product. Approximately half of all milk banks in North America do not have the resources to measure the macronutrient content of donor milk, which means pooling is done heuristically. We propose a data-driven framework combining machine learning and optimization, to predict macronutrient content of donations and then optimally combine them in pools, respectively. In collaboration with our partner milk bank, we collect a data set of milk to train our predictive models. We rigorously simulate milk bank practices to fine-tune our optimization models and evaluate operational scenarios such as changes in donation habits during the COVID-19 pandemic. Finally, we conduct a year-long trial implementation, where we observe the current nurse-led pooling practices followed by our intervention. Pools created by our approach meet clinical macronutrient targets between 31% to 76% more often than the baseline, while taking 67% less recipe creation time.

Addressing hospital emergency department (ED) overcrowding is a critical challenge for many healthcare systems worldwide. To address this challenge, many hospitals have been experimenting with innovative patient flow designs. A promising new design is to separate patients who can be served vertically (e.g., on a regular chair as opposed to horizontally on an ED bed) and route them to a different area termed the Vertical Processing unit, also known as the Rapid Medical Assessment (RMA) unit.  While this can potentially increase operational efficiency by addressing the problem of bed availability, it can degrade performance if patients are not correctly routed through the system.  Successful implementation of this design, thus, significantly depends on understanding which patients should be routed to the RMA unit. To assist our partner hospital, we developed a machine learning model trained on large-scale data capable of providing a personalized risk score for each arriving patient on whether they will eventually need an ED bed.  We then feed these risk scores to an analytical model of patient flow to characterize the optimal protocol for utilizing the RMA unit. We find that the optimal protocol depends not only on the predicted risk scores but also on the machine learning model’s accuracy and ED characteristics. Finally, we use simulation analyses to compare the performance of our recommended RMA-based design with more traditional ED flow approaches such as a “fast track” or a “physician in triage” based design. Our results suggest that following the RMA design under our recommended protocol can bring several advantages to EDs. It outperforms traditional patient flow designs due to the dynamic and efficient use of ED resources, especially in settings with a higher prevalence of high-acuity patient cases. Overall, this work provides a roadmap to healthcare systems that seek to implement data-driven patient flow systems to improve ED operations.

Behavioral health interventions, delivered through digital platforms, have the potential to significantly improve health outcomes, through education, motivation, reminders, and outreach. We study the problem of optimizing personalized interventions for patients to maximize some long-term outcome, in a setting where interventions are costly and capacity-constrained. This paper provides a model-free approach to solving this problem. We find that generic model-free approaches from the reinforcement learning literature are too data intensive for healthcare applications, while simpler bandit approaches make progress at the expense of ignoring long-term patient dynamics. We present a new algorithm we dub DecompPI that approximates one step of policy iteration. Implementing DecompPI simply consists of a prediction task from offline data, alleviating the need for online experimentation. Theoretically, we show that under a natural set of structural assumptions on patient dynamics, DecompPI surprisingly recovers at least 1/2 of the improvement possible between a naive baseline policy and the optimal policy. At the same time, DecompPI is both robust to estimation errors and interpretable. Through an empirical case study on a mobile health platform for improving treatment adherence for tuberculosis, we find that DecompPI can provide the same efficacy as the status quo with approximately half the capacity of interventions. (Joint work with V.F. Farias, J.J. Boutilier, J.O. Jonasson, E. Yoeli)