Partner(s): Indus Hospital & Health Network in Pakistan (IHHN)
Status: In progress
Team: Rayid Ghani, Jennah Gosciak, Liliana Millán, Atul Raghunathan. Extension of the project started at DSSG 2022 by our team of fellows (Olumuerjiwa Fatunde, Jennah Gosciak, Christina Last, and Kenedy Odongo)


The Indus Hospital and Health Network (IHHN)  is a free, nonprofit,  healthcare provider in Pakistan serving more than 5.4 million patients a year.  The demand for the Emergency Department (ED) services outstrips the available staff and infrastructure, and patients regularly face long wait times and overcrowding. Also, due to high patient volume and limited hospital resources, the hospital is constrained in its ability to triage and diagnose patients efficiently and effectively. We have been working with them to develop a machine-learning system that generates a candidate set of diagnosis (ICD-10) codes for each patient visit based on the notes written by nurses and physicians during the clinical workflow.


According to the World Bank, 50% of deaths and 40% of disease burden in low-and-middle-income countries are related to lack of timely and effective emergency care. Patients seeking emergency care are treated by an overburdened hospital – 98% of providers in Pakistan reported their facilities being inadequate for emergencies, and 98% of care-seekers reported not being satisfied.  A key reason is misinterpretation of symptoms and mis-triage by first-level providers, often due to lack of time and resources to create structured and consistent medical diagnosis (ICD-10) codes. Health systems in many parts of the world perform diagnosis coding primarily for billing purposes but then use it throughout the hospital workflow. IHHN, a free hospital that serves millions of patients, cannot afford to manually code patient records. Developing this system to augment the limited coding resources will allow them to provide more timely, more effective, and more equitable care.


The goal of our project is to improve IHHN’s ability to provide better care to their emergency department patients, particularly lower-income patients who rely on IHHN’s free emergency department services and often receive delayed or less effective medical care. Long waits at an emergency department increase the likelihood that patients will give up on seeking care, leading to worse medical outcomes.  We will support the medical staff in the ER by building an interactive ML system that uses the information collected during the intake and triage process  (previous and current vitals, triage notes, and other medical case notes)  to 1) generate a list of ICD-10 diagnosis codes and 2) use those standardized codes to generate recommendations for follow-up tests and procedures.  Beyond supporting real-time decision-making in the ER, these codes enable disease tracking and training new staff on evidence-based care practices.


  • Historical data from the Electronic Medical Records system in the Emergency Department provided by IHHN from Karachi’s hospital including notes of nurses and physicians
  • Priority disease areas for the hospital, as well as a criterion for narrowing this list if needed
  • Official ICD-10 codes and their descriptions from the US Centers for Medicare & Medicaid Services


The machine-learning system uses EHR data including notes from medical staff for each patient visit to generate a set of candidate ICD-10 codes. We evaluate our performance by measuring the number of actual diagnoses our system was able to capture for an average visit.


In Progress