Researchers have previously applied machine-learning-based artificial intelligence (AI) in hematologic diseases, including sickle cell disease and lymphoid and myeloid malignancies. A new study published in Blood adds differential diagnosis of bone marrow failure syndromes to this growing list.
Researchers at the National Heart, Lung, and Blood Institute (NHLBI) sought to create a tool that could help non-experts distinguish inherited bone marrow failure syndromes more easily from acquired disease using clinical and laboratory variables that are commonly collected at a patient’s first clinical visit. Fernanda Gutierrez-Rodrigues, PhD, a postdoctoral researcher at NHLBI and one of the study authors, explained, “If you can identify the patients that you think have immune pathophysiology, they can be treated with immunosuppression right away.”
Using genetic, laboratory, and clinical data, the team studied a cohort of 359 patients with bone marrow failure; patients with available genomic data were classified as having either inherited or acquired disease. They then applied a machine learning model to this cohort, entering patient data on 27 laboratory and clinical parameters. The machine learning program analyzed the respective patterns among variables and used them to develop an algorithm that could correctly predict pre-labeled inherited and acquired cases.
The researchers learned that for accurate prediction, the computer needed to learn patterns derived from a large number of cases. This was a problem for certain inherited diseases that were underrepresented in their cohort, such as Diamond-Blackfan anemia. After the model was adapted to exclude underrepresented cases, the algorithm used 25 of the clinical and laboratory variables to correctly predict acquired versus inherited disease in 89% of cases, using a validation cohort of 127 patients.
The model was better at predicting acquired disease than inherited disease in adult patients presenting with hypocellular marrow and cytopenias. The respective correct prediction was 92% versus 79%, probably because the heterogeneity among inherited syndromes and the relatively fewer number of cases made it more difficult for the AI program to recognize patterns.
The authors found that a comprehensive clinical and laboratory evaluation of all 25 variables was critical in differentiating acquired from inherited bone marrow failure. Telomere length was the top parameter; short telomere length mostly implicated an inherited telomere biology disorder, and a normal telomere length mostly indicated an immune aplastic anemia. Additional important parameters included age, sex, blood counts, history of longstanding cytopenias or macrocytosis, and mucocutaneous findings.
“When telomere length data [were] not available, we observed that patients with very severe aplastic anemia, older than 18 years, with no consanguinity, no family history, or a phenotype suggestive of inherited disease – they are very likely to have immune aplastic anemia,” Dr. Gutierrez-Rodrigues added.
One limitation is that data about paroxysmal nocturnal hemoglobinuria (PNH) clones were not included in the algorithm, even though they are an important factor in the differential. Dr. Gutierrez-Rodrigues also pointed out that the algorithm is not designed to interpret genetic reports or predict variant pathogenicity.
Practitioners can access a free app to use the algorithm through the NHLBI website (dir.nhlbi.nih.gov/DDxAA). Clinicians can input values, and the algorithm will report the likelihood of their patient having immune aplastic anemia versus inherited, or, for some conditions, whether it belongs to the patient subset that it can’t yet successfully categorize. For areas in which genetic testing isn’t available for all patients, it may help physicians prioritize the patients who truly need such testing.
Dr. Gutierrez-Rodrigues and colleagues hope to continue to refine the algorithm using data from new cases, particularly for pediatric patients, where the model’s predictive power was limited. As more data are entered, they also hope to enhance the algorithm to successfully recognize and categorize more types of inherited diseases.
Any conflicts of interest declared by the authors can be found in the original article.
Ruth Jessen Hickman, MD, is a freelance medical and science writer based in Bloomington, Indiana.
Gutierrez-Rodrigues F, Munger E, Ma X, et al. Differential diagnosis of bone marrow failure syndromes guided by machine learning [published online ahead of print, 2022 Dec 21]. Blood. doi:10.1182/blood.2022017518.