Researchers have developed, validated, and implemented an accurate and user-friendly machine-learning model for the diagnosis of heparin-induced thrombocytopenia (HIT), according to a report published in eClinicalMedicine.
The investigators conducted a prospective, multicenter, observational study of patients with suspected HIT between 2018 and 2021 and hypothesized that machine-learning algorithms could be used to develop an accurate and user-friendly diagnostic tool to integrate diverse clinical and laboratory information.
The team collected detailed clinical information and laboratory data, including results from immunoassays (chemiluminescent immunoassay [CLIA], particle-gel immunoassay [PaGIA], and enzyme-linked immunosorbent assay [ELISA]) of participants. They used the washed platelet heparin-induced platelet activation assay (HIPA) as the reference standard and evaluated 5 separate machine-learning prediction models for each immunoassay.
A total of 1393 patients with suspected HIT (median age, 67.02 years; interquartile range, 57.4-75.1; 63.8% men and 36.2% women) were included in the study. Of those, 119 patients were HIPA positive and 1274 patients were HIPA negative, translating to a HIPA-diagnosed HIT prevalence of 8.5%.
To develop a practical prediction model intended for bedside use, the researchers conducted a step-wise feature selection process, balancing trade-offs between model accuracy and usability. In the feature selection process, they used a training dataset comprising 75% of the patient cohort and identified immunoassay test result, platelet nadir, unfractionated heparin use, C-reactive protein, timing of thrombocytopenia, and other causes of thrombocytopenia as predictor variables.
The researchers found the best performing models were support vector machine algorithms for CLIA and ELISA and a gradient boosting machine algorithm for PaGIA. In the validation dataset, comprising 25% of patients, they reported the area under the receiver-operating characteristic curve (AUC) of each of these models was 0.99.
The team compared numbers of false-negative and false-positive patients of these algorithms with those of the currently recommended diagnostic algorithm (4Ts score, immunoassay). They found the numbers of false-negative patients were reduced by 50% for ELISA, 66.7% for PaGIA, and 64.3% for CLIA. Also, the numbers of false-positive individuals were reduced by 29.8% for ELISA and 68.5% for PaGIA and increased by 29.0% for CLIA.
Collectively, these models underlie the novel multivariable diagnostic prediction tool, the TORADI-HIT algorithm. To facilitate its use at the bedside, the researchers made it accessible online.
“The TORADI-HIT algorithm has the potential to reduce delayed diagnosis and overtreatment in clinical practice. Future studies shall assess usability and performance in other patient populations and health care systems,” the researchers concluded in their report.
The primary limitations of the study were inclusion of mostly patients from Switzerland and use of results only from commonly employed immunoassays for model development and validation.
Disclosure: Some study authors declared affiliations with biotech, pharmaceutical, or device companies. Please see the original reference for a full list of authors’ disclosures.
Nilius H, Cuker A, Haug S, et al. A machine-learning model for reducing misdiagnosis in heparin-induced thrombocytopenia: A prospective, multicenter, observational study. eClinicalMedicine. 2022;55:101745. doi:10.1016/j.eclinm.2022.101745