Pioneering AI-based approach identifies rare disease patients using electronic health records

Rare diseases (RDs) are defined as those that occur in one for every 2,000 people in the European Union and less than 200,000 people in the United States each year. To date, about 7,000 RDs have been identified; however, the diagnosis of RDs remains a challenge.

A new Frontiers in Neurology study describes the outcomes of an automated artificial intelligence (AI)-based methodology to diagnose Pompe disease, a known RD.

Study: An artificial intelligence-based approach for identifying rare disease patients using retrospective electronic health records applied for Pompe disease. Image Credit: SOMKID THONGDEE / Shutterstock.com

Background

Sometimes an isolated symptom of a rare disease is misdiagnosed as a common disease, which leads to “premature closure.” Other times, a prolonged diagnostic journey that includes consultations with various physicians and a subsequent plethora of examinations and treatments can impair diagnostic judgment. On average, it takes about eight years, with some reporting up to 28 years, to diagnose RDs.

Typically, it takes between 2.5 months and 144 months to correctly diagnose Pompe disease. Estimation methods based on genetic databases have indicated that Pompe disease prevails in one in every 23,232 individuals. However, survey-based investigations, which are based on clinical centers that provide Pompe disease treatment, revealed the prevalence rate of this disease to be 1 in every 350,914 persons.

A significant variation in the prevalence of this disease has been observed geographically. For example, the highest and lowest prevalence of Pompe disease was found among East Asian and Finnish people, respectively.

Pompe disease has been divided into two groups based on the phenotype. These include infantile-onset Pompe disease (IOPD) and late-onset Pompe disease (LOPD).

Electronic Health Records (EHRs) have a comprehensive collection of healthcare data. Unfortunately, this data is extremely heterogenous and often incomplete, thereby causing difficulties for traditional automated diagnostic methods.

Currently, AI-based methods are being designed to extract relevant information from unstructured medical data. This type of automated screening can be used for large-scale rare disease screening, which could increase the rate of diagnosing RDs in a cost-effective manner.

About the study

The current study used an automated AI-based approach developed by Symptoma to detect Pompe disease based on available retrospective anonymized EHRs.

A patient with Pompe disease detected by AI was referred to as “flagged.” The data of a flagged patient was divided into their patient profile, clinical presentation, and hidden disease patterns.

Once AI determined that adequate evidence was available to support the diagnosis of Pompe disease, it flagged the identified patient for further assessment. This patient was evaluated by generalist physicians (GP), with some subsequently recommended to specialist physicians (SP) for Pompe disease.

Study findings

A total of 104 patients were flagged by AI out of 350,116 patients who were admitted to any of the five Landeskliniken Salzburg clinic groups through the one-time application.

AI detected one possible patient with Pompe disease out of every 3,366 patients screened. These patients were reviewed by GPs, 22 patients of whom were identified for further evaluation by SPs. GPs classified these patients into three groups, including diagnoses, suspected, and reduced suspicion.

All patients presented to SPs were further evaluated and grouped as definite (five patients), probable (two patients), possible (six patients), inconclusive (six patients), and unlikely (three patients).

About four possible Pompe patients for each clinic were identified every month. On average, seventeen patients were examined to identify one suspected patient.

The AI-based automated screening method, including the review round with GPs and SPs, took less than one month to detect 21 suspected patients for every clinic. Therefore, this approach required 5.47 patients to be reviewed manually to identify one suspected patient.

Free-text documentation was used for patient characteristics analysis. Fatigue, lower back pain, and headache were the most common symptoms.

Some of the features related to organ abnormalities included hepatomegaly and splenomegaly. Symptoms related to muscular impairment were found to be important for differentiating patients suspected to be suffering from Pompe disease.

AI identified characteristic features of Pompe disease based on EHRs, including myopathy, muscle hypotony, muscle weakness, scapula alata, and myalgia. However, clinical features can vary, particularly for LOPD.

Fatigue was inversely related to cardiomegaly, which is typically found in IOPD. However, some features are less clinically significant in newborns as compared to older patients. 

Conclusions

Like any other RD study, the current study had a small sample size that limited the potential to obtain statistically significant findings. However, the AI-based methodology enabled the analysis of a large dataset that allowed the researchers to obtain statistically significant results.

Another limitation of the study is that AI could only access documented evidence or those present in EHRs. However, it is not uncommon for all relevant evidence to be inadequately documented or the doctor’s impression not recorded, which can generate biased findings.

Despite the limitations, the current study revealed how AI could be used to diagnose Pompe disease using retrospective EHRs. This method can improve the timing and accuracy of diagnosis of other RDs as well.