Abstract: SA-PO012
Use of Natural Language Processing and Deep Learning to Analyze Kidney Ultrasound Reports and Their Correlation with CKD Diagnosis
Session Information
- Augmented Intelligence, Large Language Models, and Digital Health
October 26, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Koraishy, Farrukh M., Stony Brook University, Stony Brook, New York, United States
- Wang, Chenlu, Stony Brook University, Stony Brook, New York, United States
- Banerjee, Ritwik, Stony Brook University, Stony Brook, New York, United States
- Kuperstein, Harry, Stony Brook University, Stony Brook, New York, United States
- Malick, Hamza, Stony Brook University, Stony Brook, New York, United States
- Tahir, Hira, Stony Brook University, Stony Brook, New York, United States
- Bano, Ruqiyya, Stony Brook University, Stony Brook, New York, United States
- Sakhuja, Priyal, Stony Brook University, Stony Brook, New York, United States
- Hajagos, Janos G., Stony Brook University, Stony Brook, New York, United States
Background
Natural language processing (NLP) can analyze unstructured data in imaging reports for clinical correlation, but this has not been reported in the context of kidney ultrasound reports and chronic kidney disease (CKD).
Methods
From the set of 1,068 patient ultrasound reports, NLP models were developed for kidney echogenicity and length. Kidney echogenicity was divided into ‘normal’, ‘echogenic’ or ‘others’ categories using NLP Toolkit to isolate sentences pertinent to echogenicity and Parse Tree Method and BioBERT to determine the word-level and sentence-level echogenicity classification respectively. Embeddings from Language Models was used to extract kidney size data and length classified as ‘normal’ or ‘small’ based on percentiles.100 reports were randomly selected for annotation by nephrologists to develop ground-truth ultrasound labels that were initially tested on the model. Subsequently the model was used to analyze the association of kidney ultrasound features with CKD diagnosis using logistic regression (LR) models.
Results
The word-level NLP method (Figure 1) demonstrated higher accuracy and precision for the classification of increased echogenicity compared to the sentence-level method. Subsequently only the word -level NLP model was used for further analyses with clinical correlations. Based on the 10% percentile, kidneys measuring under 8.5 cm in females and under 9.0 cm in males were categorized as “small”. On multivariable LR, in addition to traditional factors like age, sex, diabetes, hypertension, heart failure and AKI; the presence of bilaterally echogenic kidneys was a strong predictor of CKD (OR = 5.49 [3.44, 8.75]; p<0.0001). The presence of bilaterally small kidneys was also a significant predictor (OR = 3.77 [1.25, 13.87)]; p=0.046).
Conclusion
Advanced NLP models with detailed text analysis can accurately detect CKD features in kidney ultrasound reports.