�԰�����

By Dave DeFusco

At a time when obesity and related diseases like diabetes are on the rise, being able to predict weight gain and changes in body shape could help prevent serious health problems before they start. That’s the promise behind a new kind of artificial intelligence developed by Dengyi Liu, a student in the Katz School’s Ph.D. in Mathematics.

Liu, together with Dr. Honggang Wang, chair of the Katz School’s Department of Graduate Computer Science and Engineering, and Dr. Hua Fang, a professor of computer and information science at the University of Massachusetts (UMass) Dartmouth and UMass Chan Medical School, presented their work at the 47th IEEE Engineering in Medicine and Biology Conference in Copenhagen, Denmark. Their tool, called a Fuzzy Deep Neural Network (Fuzzy DNN), is an advanced kind of AI designed to make health predictions that are more accurate and more reliable than previous methods.

“Health data is messy and uncertain,” said Liu. “People forget what they ate, measurements can be imprecise and human bodies respond in very different ways. We designed the Fuzzy DNN to make sense of all that uncertainty.”

If you’ve ever tried to lose weight or track your health, you probably know how unpredictable it can be. Maybe you eat the same thing two days in a row but gain weight one day and not the other. That inconsistency is common in large health studies, too, where data can come from self-reported food diaries, different lab tests and people with very different lifestyles.

Traditional AI models, like deep neural networks, are powerful tools for finding patterns in complex data, but they don’t handle uncertainty very well. They assume the data is clean and consistent, which is rarely true in real-world health research.

Fuzzy logic is a kind of computing that deals with gray areas instead of black-and-white answers. Rather than saying something is either true or false, like “you are overweight” or “you are not,” fuzzy logic allows for degrees of truth, like “you are somewhat overweight.”

The researchers created a hybrid system that blends traditional deep learning with fuzzy logic. Here’s how it works:

• The MLP (Multilayer Perceptron) Network is the deep learning part. It looks at the big picture in the data and predicts general trends in weight, body mass index (BMI) and waist size.
• The Fuzzy Correction Module is like a fine-tuning tool. It adjusts the predictions based on fuzzy rules, taking into account uncertainty or inconsistencies in the data.

This combination allows the model to make better predictions even when the data is noisy or incomplete.

“It’s like having two doctors look at your health data,” said Liu. “One looks at long-term patterns across many people, and the other looks at your unique details and makes small corrections based on experience and fuzzy knowledge.”

The team used a large dataset from the National Institutes of Health, known as the AHEI (Alternate Healthy Eating Index), which contains about 10,000 records. These included measurements of BMI, weight and waist size, along with diet scores and other health indicators like blood pressure and cholesterol.

Before feeding the data into the AI model, they cleaned it up using a method called multiple imputation, which fills in missing values with estimated ones. They also normalized the data, making sure every measurement was on the same scale, so the AI could process it fairly. Then, they trained and tested their model using several versions of the data to see how different factors, like gender or whether someone was in a control group, affected prediction accuracy.

The Fuzzy DNN crushed the competition. It performed dramatically better than three commonly used methods: Linear Regression, a basic statistical model; LSTM, a deep learning model for time-based data; and Standard Deep Neural Networks, without the fuzzy logic.

“The Fuzzy DNN was the most precise and most stable, even with incomplete or noisy data,” said Liu. “That makes it a powerful tool for real-world health predictions.”

Predicting changes in body weight, BMI and waist size may sound simple, but it’s a key part of preventing serious conditions like diabetes, heart disease and metabolic syndrome.

“This research brings us closer to personalized, preventative healthcare,” said Dr. Wang. “By identifying subtle trends early, doctors can intervene before a problem becomes serious.”

Dr. Fang said incorporating fuzzy logic into deep learning is a promising direction for healthcare AI. “It makes the model more human-like in how it deals with uncertainty," she said, "more like how a clinician thinks.”

The team plans to expand the model to predict more serious health outcomes, such as Type 2 diabetes or heart disease, and to make the model more transparent so that doctors and patients can understand why it makes certain predictions.

“We believe this work can lay the foundation for smarter, fairer and more personalized health tools,” said Liu. “It’s just the beginning.”