Tech Meets Philanthropy: Using Deep Learning to Predict Medical Emergency Recurrence

Few things inspire gratitude as intensely as having someone save the life of your parent. So after her mother had life-changing surgery in Israel recently, Intuit’s Sigalit Bechler was fueled with a desire to give back. Specifically, Bechler wanted to contribute to medical research at the facility that helped her mother, Sheba Medical Center –

Few things inspire gratitude as intensely as having someone save the life of your parent.

So after her mother had life-changing surgery in Israel recently, Intuit’s Sigalit Bechler was fueled with a desire to give back. Specifically, Bechler wanted to contribute to medical research at the facility that helped her mother, Sheba Medical Center – the biggest hospital in the Middle East.

And so began a first-of-its-kind collaboration between Intuit and the hospital. Bechler, a data scientist in Israel, reached out to her innovation team and to tech-savvy doctors at the hospital. The doctors suggested several projects that would fill unmet needs at the hospital, and the innovation team decided to devote their “We Care and Give Back” time to finding solutions.

“I felt fortunate to be part of a company that really enables such a massive contribution to the community,” Bechler says. “I also felt grateful to have the opportunity to be part of this group of great researchers. It was an extremely wonderful feeling to turn from being passive to being active, both regarding my mother’s situation and helping other people.”

Tech meets philanthropy

The collaboration led to the development of three models centering on patient traffic through the extremely busy emergency room (ER). The models used machine learning (ML) and Natural Language Processing (NLP) to analyze the data of a half-million patients who visited the ER. The models focused on predicting the mortality of arriving patients, cataloging radiologists’ interpretation of CAT scans, and – most importantly – predicting return trips back to the ER.

Predicting the mortality of patients arriving at the ER

This is critical for prioritizing treatment in the ER (acute patients get priority). Currently, an ER nurse scores the acuteness of a patient’s condition on a 5-point scale based on limited data.

The Intuit team’s model took far more data into account – including triglyceride level, demographic data and vital measures – to determine which patients were in a life-threatening condition. As a result, the Intuit team was able to predict who would pass away within two days of arriving in the ER with a much higher rate of accuracy than previously (0.8 AUC, a measure of an algorithm’s performance).

Automatic cataloging of radiologists’ interpretation of CAT scans

Radiologists interpret CAT scans in natural language without specifying if there is an acute condition. This interpretation is sent to the patient/treating doctor (the doctor may not be staff at the hospital where the imaging was done). The Intuit team used NLP algorithms to transform medical text in Hebrew into structured information that could be quantified as acute, chronic or nothing wrong. This was a challenge because available modules for NLP are usually for English tex.

The team achieved 90 percent accuracy in their severity predictions — outperforming the 85 percent accuracy benchmark the doctors provided using data in English.

Predicting patients returning to the ER

About 13 percent of patients return to the ER after discharge. Predicting who those patients will be matters because in some cases (e.g., if a patient returns within 30 days) the hospital will have to foot the bill.

So after examining 170 different parameters registered during a patient’s stay to predict outcomes for released patients (including re-hospitalization within 30 days), the Intuit team achieved an AUC accuracy score of 0.7.

“There were a lot of challenges,” says Bechler. “There are a lot of issues with the data — most of it is very confidential and sensitive. But we found a solution to that and it turned out to be amazing.”

Using deep learning to predict breast cancer recurrence

ER visit predictions aren’t the only medical encounter deep tech can help improve. At the recent Grace Hopper Celebration, Intuit’s Noah Eyal Altman unveiled her research showing how machine learning can improve breast cancer recurrence prediction.

Chemotherapy does indeed reduce recurrence and mortality for early stage breast cancer; without it, almost half of the women over all breast cancer subtypes will recur. But it also brings with it a great cost to patients and caregivers alike – both economically and emotionally.

Therefore, estimating recurrence risk is clinically important; it allows caregivers the opportunity to offer adjuvant therapy only to patients at high risk. This is done by assessing individual risk through measuring gene expression froףm the primary tumor.

For this study, the team identified a dataset of 1,519 women with early-stage breast cancer for which whole transcriptome analysis was done on the primary tumor and a long-term follow-up was available. They showed that coupling deep learning and random forest can achieve higher accuracies than ever reported in a large population (84.79%) and that this accuracy is higher than obtained with each algorithm alone.

Long story short – deep learning can be used in biomedical research even with training sets of moderate size, and in a supervised setting, should be further explored as a dimensionality reduction method in biomedical research.

A deeply satisfying experience

Thirteen Intuit data scientists in four teams took part in the Aug. 5 – 7 hackathon. Those who participated said they found the experience fulfilling and satisfying, and were excited to be able to demonstrate Intuit’s WCGB value. The partner doctors were amazed at how fast the team was able to achieve meaningful results, and said they plan to publish papers about them.

“Every year the population is getting older so emergency rooms need to deal with more patients,” says Dr. Eyal Zimlichman, deputy director of Sheba hospital. “That’s why we need to find out-of-the-box solutions to help make the process and results better in scale. We wanted to provide the Intuit team with a GitHub equivalent for medical knowledge and resources. That’s something that we tried to spark in this hackathon, and this is why this pilot is so important.”

“I was surprised to see how the team members felt connected right from the start to the project and instantly asked to join it,” says Shimon Shahar, a distinguished data scientist at Intuit Israel. “Most of our efforts were dedicated to analyze and help triage patients in the emergency room. And although many companies have healthcare divisions, we hardly see this type of unique activity between tech companies and hospitals.”

As for Bechler, she says she hopes to make the hackathon an annual event. “I hope that this will be a catalyst for other companies to get involved in philanthropic technical work,” she says. “Today more than ever, I believe that collaborations with a scientific interest can significantly advance medicine.”