Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2062
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-21T01:09:58.715766Z
status: published
lens: g_space
sector: health
headline: Construction and Validation of a Predictive Model for Voriconazole Treatment Efficacy in Influenza-Associated Pulmonary Aspergillosis Based on Machine Learning
dek: Purpose To analyze risk factors affecting voriconazole efficacy in influenza-associated pulmonary aspergillosis (IAPA) and build interpretable machine learning (ML) predictive models. Patients and methods This retrospective study enrolled voriconazole-treated IAPA patients into a primary retrospective set (Jan 1, 2020-Jul 31, 2025) and a separate temporal external validation set (Aug 1, 2025-Apr 30, 2026). To avoid data leakage, the primary set was stratified by treatment response and randomly split at a 7:3 rat…
gain_title: A CatBoost model flagged likely voriconazole non-response in IAPA patients with AUC 0.88 internally and 0.83 on a later temporal validation set, which could support individualized antifungal therapy.
problem_title: (none)
trace_subject: (none)
gain_reading: A CatBoost model flagged likely voriconazole non-response in IAPA patients with AUC 0.88 internally and 0.83 on a later temporal validation set, which could support individualized antifungal therapy.
gain_evidence: The CatBoost model achieved the optimal discrimination, with internal-validation AUC = 0.88 | temporal external-validation AUC = 0.83 | guide individualized voriconazole therapy for IAPA
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers analyzed patients with influenza-associated pulmonary aspergillosis who were treated with voriconazole. They built seven machine learning models from baseline clinical variables to predict whether treatment would work, using a primary retrospective set and a later temporal validation set.

The work matters because it points toward tailoring antifungal treatment to patient risk, rather than treating all cases alike. It remains a retrospective single-source analysis. The external validation interval was wide, and the machine learning feature ranking only partly agreed with logistic regression.
limitation: The SHAP-based feature importance only partly matched the logistic regression ranking, and the external validation AUC had a wide confidence interval (0.70-0.94), indicating uncertainty about performance in new patients.
tag: Evidence-backed gain
key_points: Internal set outcomes: 284 voriconazole-responsive patients and 56 non-responsive patients, defined by clinical remission, radiological absorption and inflammatory biomarker reduction. | Logistic regression identified four independent predictors of poor response: ICU admission, concomitant midazolam use, higher baseline neutrophil percentage and lower baseline platelet count. | CatBoost performed best, with AUC 0.88 internally and 0.83 (95% CI 0.70-0.94) in temporal external validation. | Decision curve analysis showed positive net clinical benefit for all seven models across threshold probabilities of 0.00-0.90.
rundown: The study enrolled voriconazole-treated patients into a primary retrospective set covering January 2020 to July 2025 and a temporal external validation set covering August 2025 to April 2026. The primary set was stratified by response and split 7:3 into training and internal validation sets before feature screening, to avoid data leakage.

From 42 candidate baseline variables, screening within the training data only (univariate tests, 10-fold LASSO and variance inflation factor filtering) retained 11 non-collinear predictors. Seven machine learning algorithms were then built and compared.

SHAP analysis was used to interpret the best model and highlighted ICU admission and midazolam as important contributors. Its ranking differed in part from the multivariate logistic regression ranking.
sources:
- peer_reviewed | Drug Design, Development and Therapy | https://doi.org/10.2147/dddt.s641070 | 2026-10-12
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

2cede59c7a81190ff370e6f4c8a3cb2d82864d419cd8c3b80e3fc3e68dfccd12

previous

0000000000000000000000000000000000000000000000000000000000000000