Public Health

Problem

A public health programme usually knows its budget before it has a model: a fixed number of home visits, a fixed number of hospital beds, a fixed amount of a radiologist's time. A model helps only if it orders the cases well inside that budget, and only if the quantity it is scored against is the quantity that matters.

Approach and findings

At Wadhwani AI, with India's Central TB Division and the National Institute for Research in Tuberculosis, we built models that predict which patients will interrupt treatment for 30 days or more. Health workers can reach about a fifth of patients, so the models were trained and judged on recall within that fifth rather than on overall accuracy. In a six-month evaluation, targeting the top fifth with the model would have reached 5,587 patients who went on to interrupt treatment, against 2,974 for the best rule-based baseline and 1,808 for random targeting. The work received the Outstanding Paper Award at ML4H 2022, and the models were taken up against the national register of about 20 million patients.

For COVID-19, we posed the choice of a lockdown schedule as a reinforcement learning problem over an epidemic model fitted to Mumbai's case data. The epidemic's parameters were themselves uncertain, so a second agent chose the worst parameters during training. The resulting policy had the best worst-case score on both objectives and a lower average score than policies trained on fixed or sampled parameters, which is the cost of the robustness.

In mammography, a clinician's outline encloses the lesion together with a margin of healthy tissue, so a segmentation model scored against the outline is graded on the annotation and not on the lesion. We introduced two metrics, intersection over prediction and intersection over annotation, that keep the two apart. Our model reached an intersection over prediction of 0.90, against 0.66 for U-Net.

Papers

Predicting Treatment Adherence of Tuberculosis Patients at Scale
Mihir Kulkarni*, Satvik Golechha*, Rishi Raj Grandhe*, Jithin K. Sreedharan*, Ankit Bhardwaj, Santanu Rathod, Bhavin Vadera, Jayakrishna Kurada, Kirankumar Rade, Sanjay Mattoo, Rajendra Joshi, Alpan Raval
Machine Learning for Health Symposium (ML4H), 2022. Outstanding Paper Award. [Paper]

Robust Lock-Down Optimization for COVID-19 Policy Guidance
Ankit Bhardwaj*, Han-Ching Ou*, Haipeng Chen, Shahin Jabbari, Milind Tambe, Rahul Panicker, Alpan Raval
AAAI Fall Symposium on AI for Social Good, 2020. [Paper] [Video]

High Precision Mammography Lesion Identification From Imprecise Medical Annotations
Ulzee An, Ankit Bhardwaj, Khader Shameer, Lakshminarayanan Subramanian
Frontiers in Big Data, 2021. [Paper]

Collaborators

• Alpan Raval (Wadhwani AI)

• Milind Tambe (Harvard University)

• Shameer Khader (Sanofi)

• Lakshminarayanan Subramanian (NYU, advisor)