Machine-learning approaches show promise for use in identifying potential donor kidneys at high risk of organ nonuse or nonrecovery, according to a study in JAMA Surgery.
Using information from the United Network for Organ Sharing (UNOS), the researchers evaluated the use of artificial intelligence (AI) approaches to make predictions about the use versus nonuse of potential donor kidneys. The study evaluated machine learning (ML) models using structured data on donor characteristics, as well as natural language processing (NLP) models using unstructured, free-text donor narratives. The free-text data included comments from the UNOS admission course, medical and social history, as well as donor highlights.
The AI approaches were evaluated for their ability to classify donors regardless of recovery status versus those who had at least one kidney recovered for transplant. Performance was compared with that of a model using the Kidney Donor Profile Index (KDPI). A training and validation cohort consisted of 9555 donors offered to the study center between 2015 and 2020; a test cohort comprised 2481 donors from 2021.
Just 20% to 30% of potential donors had at least one kidney transplanted. The model using the KDPI had an area under the receiver operating characteristic curve of 0.69, with accuracy of 0.64. Performance was almost identical for two multivariable ML models based on structured donor data (logistic regression and random forest classifier models).
A classic “bag of words” NLP model showed the best performance with the random forest classifier: area under the curve, 0.70 and accuracy, 0.59. An advanced Bidirectional Encoder Representations from Transformers model met this level of performance only after the addition of basic donor information.
Models using free text were “slightly inferior” to models using structured data. Analysis of feature importance and Shapley additive explanation summaries provided information on conditions potentially affecting donor selection: Terms implying chronic disease tended to have negative effects, whereas terms implying trauma appeared positive.
The findings suggest that ML models can potentially predict donors with high-risk kidneys that are ultimately not used for kidney transplant. The researchers conclude: “The use of structured data is likely to expand the possibilities, but further exploration of new approaches…will be necessary to develop explainable models with high predictive metrics” [Sageshima J, et al. Prediction of high-risk donors for kidney discard and nonrecovery using structured donor characteristics and unstructured donor narratives. JAMA Surg, published online November 1, 2023. doi: 10.1001/jamasurg.2023.4679].
