Diagnosing disease with AI could be the new norm in personalized medicine

A new AI diagnostic tool uses microbiome data and lifestyle factors to predict risk of multiple diseases, ushering in a new era of personalized healthcare.

Our bodies are brimming with trillions of microorganisms, including bacteria, fungi, parasites, and viruses. Collectively, these living entities make up our microbiome, which is shaped by our DNA, external environment, and diet.

The microbiome’s composition is unique to each individual and not only  gives information about our current health status — even our emotional well-being — but also how well we age and the likelihood of developing chronic disease later in life.   

Using artificial intelligence (AI), scientists have been able to correlate the composition of an individual’s gut with certain diseases, taking personalized medicine to the next level.

For example, last year, researchers introduced a microbiome-based predictive model that could accurately and non-invasively predict type 2 diabetes and inflammatory bowel disease (IBD), two ubiquitous non-communicable diseases. But this model did not take the possibility of comorbidities into account, limiting its practicality.

With the global shift toward population aging, a growing number of people are living with more than one chronic disease, a scenario often overlooked by healthcare practitioners.  

To confront this reality, tools that facilitate the detection of multiple diseases simultaneously are needed, and a new AI-powered tool called Meta-Spec could do just that.     

A personalized tool to assess health status

Meta-Spec, an AI-based diagnostic tool developed by researchers in China and the United States, considers multiple factors that might contribute to disease rather than relying solely on microbiome data. The result is a more nuanced, comprehensive approach to disease detection and prediction than what other models can currently offer.  

Meta-Spec incorporates easily collected physical and lifestyle data such as diet, body mass index, and age. These details are part of a host’s “phenotype” — the collection of observable traits that stem from that person’s genes and their environment.   

“By harnessing the power of deep learning and integrating it with microbiome data, Meta-Spec offers a glimpse into a future where healthcare is more personalized, more accurate, and ultimately, more effective,” stated Xiaoquan Su, a bioinformatics professor at Qingdao University’s College of Computer Science and Technology and one of Meta-Spec’s developers.

This multi-faceted approach significantly improves the model’s disease-screening accuracy to the point of being able to simultaneously detect multiple diseases, if present.

“In the past, the focus has largely been on detecting individual diseases, often overlooking the complex interplay of various factors that influence our health,” said Shunyao Wu, a computer science professor at Qingdao University and member of Meta-Spec’s development team.

Training Meta-Spec to detect diseases

Meta-Spec classifies diseases using multitask deep learning, a machine-learning technique that uses artificial neural networks to recognize patterns. Through several different “layers”, the model learns these patterns in a given dataset by merging different microbial features with questionnaire-based host data (“metadata”).

This can range from how often the host has vivid dreams to their bowel movement quality. From this data, the model learns to associate a distinct microbiome pattern with a particular disease, and each disease’s probability is calculated.

To train and validate their model, the researchers used data from several open-source platforms that collect human microbiome specimens from volunteers, including the American Gut Project and the Guangdong Gut Microbiome Project.

In the first dataset, each patient had been diagnosed with one or more diseases, including autoimmune disease, lung disease, thyroid disorder, cancer, IBD, cardiovascular disease, and autism spectrum disorder. Patients in the second dataset had metabolic syndrome, gastritis, type 2 diabetes, and/or gout.

The keys to Meta-Spec’s accuracy

For each dataset, Meta-Spec predicted the diagnosed disease(s) more accurately than traditional machine-learning methods. When only microbiome data was factored in, both Meta-Spec and models based on traditional methods performed much worse, demonstrating that the additional information about a person’s lifestyle — including seemingly unrelated details such as how often they floss their teeth — greatly improved the model’s predictive performance.

To further test their model’s capabilities, the researchers divided the two datasets into a single-disease group and a comorbidity group with one or more additional diseases. In this case, the model also outperformed other machine-learning methods that focus only on the target disease.

Meta-Spec’s higher level of accuracy can also be attributed to its ability to rank information. By determining how much a given microbiome or phenotypical characteristic contributes to developing a certain disease and assigning a value to it, the model can make important associations.

For example, age was determined to be the most important factor in detecting cardiovascular disease, where older people are more susceptible. Interestingly, through this ranking feature, the model also linked artificial sweeteners, seafood consumption, and constipation to cardiovascular disease.

Despite its merits, the model also has limitations. “To maintain a high detection performance, Meta-Spec needs a large volume of microbiome data and host metadata. This is a common challenge for deep learning-based approaches,” Su stressed.

To help offset potential data scarcity, Meta-Spec’s developers created a hybrid model that merged data from US- and UK-based cohorts. They found that this cross-cohort approach improved the performance of the model that was trained on local data alone. Since a person’s gut microbiome depends on where they live in the world, a multicohort model could help bridge the geographical gap.   

According to Meta-Spec’s developers, once the tool is well-trained by experts, it could become a routine part of a doctor’s office visit in the not-too-distant future.  

“We hope Meta-Spec can be applied in hospitals or physical examination centers for early-stage disease prediction,” Su told us. “Meanwhile, Meta-Spec can also help microbiome scientists in the study of host–microbe and microbe–microbe interactions among multiple diseases.”

Eventually, Meta-Spec may even be available as a user-friendly app. “Users could upload the query microbiome data from any computer, and results can also be displayed in a very easy-to-understand way –– just like ChatGPT,” said Su.