Machine learning could help ID those with need for FAP genetic test

Model aware of disease's 'red flags' may aid faster diagnosis, therapy start

Margarida Maia, PhD avatar

by Margarida Maia, PhD |

Share this article:

Share article via email
An illustration of a health professional working at a computer while a person seated alongside the desk speaks.

Machine learning, a form of artificial intelligence, may help doctors to identify people suspected of familial amyloid polyneuropathy (FAP) and who should undergo a genetic test for the disease, a study in Italy suggested.

Machine learning uses algorithms to analyze data, learn from its analyses, and make a prediction based on this information.

Heart damage or cardiomyopathy, unexplained weight loss, and gastrointestinal problems strongly associate with a genetic diagnosis of [FAP], indicating “these symptoms might represent the most sensitive red flags” in diagnosing FAP, the researchers wrote.

The study, “Machine Learning for Early Diagnosis of ATTRv Amyloidosis in Non-Endemic Areas: A Multicenter Study from Italy,” was published in Brain Sciences.

Recommended Reading
An illustration of bells with the word

Intellia prepping for pivotal trials of gene-editing therapy NTLA-2001

In absence of genetic testing for FAP, diagnostic delays can be common

FAP, also known as hereditary transthyretin amyloidosis with polyneuropathy, is an adult-onset progressive disease caused by mutations in the TTR gene.

This results in the accumulation of toxic clumps of the transthyretin protein in tissues, particularly in peripheral nerves — those found outside the brain and spinal cord — leading to nerve damage or polyneuropathy. Transthyretin aggregates also can build in the heart, affecting its function (cardiomyopathy).

FAP is characterized by a wide range of disease symptoms that vary greatly between patients and are similar to other diseases, making a diagnosis “very difficult and, in most cases, delayed,” the researchers wrote. “Nowadays, several treatment options are available; thus, avoiding misdiagnosis is crucial to starting therapy in early disease stages.”

The use of machine learning “in the genetic screening for [FAP] might lead to a higher sensitive and specific diagnostic approach, thus contributing to a significant reduction in the diagnostic delay of [FAP] in non-endemic areas, as well as ensuring the early treatment for this rare inherited disease,” they added.

To test this hypothesis, scientists in Italy evaluated whether machine learning could help discriminate between patients with and without FAP, thereby helping to identify those who should undergo genetic testing for FAP.

“We aim to develop a simple guide for genetic testing that may be useful for clinicians,” they wrote.

They looked at information covering 397 adults with chronic polyneuropathy and at least one “red flag” that raised a suspicion of FAP. All had undergone a genetic test for the disease at a neuromuscular center in Palermo, Messina, Naples, or Rome.

After excluding first-degree family members, the analysis included 93 patients (median age of 68; 77% men) who tested positive for FAP-causing mutations, and 96 age- and sex-matched people who tested negative.

Among mutation-positive cases, the most common symptoms were carpal tunnel syndrome on both hands and problems in the autonomic nervous system (51% for each symptom), followed by ataxia or loss of coordination (48%), unexplained weight loss (45%), and cardiomyopathy (42%).

The autonomic nervous system is responsible for the control of involuntary bodily functions, such as heart rate, blood flow, and gastrointestinal and bladder function.

To understand how well machine learning could distinguish between polyneuropathy patients showing positive for relevant genetic mutations and those who were negative, the researchers trained a machine learning algorithm called XGBoost (XGB) that uses several “decision trees,” constructed sequentially, to create a final model.

Good sensitivity, specificity in identifying mutation-positive patients

The model distinguished positive and negative test result patients with an accuracy of 70.7%, a sensitivity of 71.2% (true-positive rate), and a specificity of 70.4% (true-negative rate).

“The results suggest the capability of the XGB model to fairly identify both positive and negative samples,” the researchers wrote, adding that the model outperformed other standard models.

Researchers then used another artificial intelligence algorithm to interpret the model’s findings and understand which factors are most important in determining who should be advised to take a genetic test for FAP.

They found that ataxia, unexplained weight loss, gastrointestinal symptoms, and cardiomyopathy were strongly associated with a positive result on the genetic test. In turn, autoimmune disease, eye involvement, diabetes, kidney symptoms, and signs of narrowing of the lower spinal canal were associated with a negative result.

While bilateral carpal tunnel syndrome and autonomic dysfunction were the most common symptoms of FAP patients, these symptoms also were found at similar rates among those with a negative test, being “similarly distributed in both screening-positive and negative patients,” the team wrote.

“Our data support the use of [AI] algorithms in clinical screening to raise the suspicion of [FAP], thus contributing to a potential reduction in the diagnostic delay in non-endemic areas,” the researchers wrote.

FAP “should be suspected if progressive [polyneuropathy] is observed in combination with ataxia, gastrointestinal problems, unexplained weight loss, and cardiomyopathy,” they added, recommending further study to “explore the clinical application of [a machine learning] algorithm” in an early diagnosis.