Scholarlink

Purpose <p>Older adults undergoing surgery are at high risk for adverse outcomes due to age-related vulnerabilities. Comprehensive geriatric assessment (CGA) can improve discharge planning but is often limited by resource constraints. We evaluated whether machine learning (ML) can support individualized discharge destination recommendations for geriatric surgical inpatients.</p> Methods <p>We developed and internally validated an AdaBoost classifier using electronic health record and CGA data from 169 patients aged ≥ 70&#xa0;years undergoing surgery at a single center. The model predicted discharge destinations (home, acute geriatric care unit, rehabilitation facility, nursing home) and was evaluated using accuracy, receiver operating characteristic (ROC) curves, and calibration within a fivefold cross-validation framework. Model performance was also compared with previously reported standard-of-care discharge decisions.</p> Results <p>The study population consisted of 169 older adults, with mean age of 80.5&#xa0;years (SD ± 6.3). The ML model achieved a micro-averaged ROC area under the curve of 0.94 [95% CI 0.92–0.96] and an accuracy of 82% [95% CI 0.76–0.86], representing a significant improvement over standard-of-care decisions (138 vs 124 correct, <i>p</i> = 0.041). Mimicking a human-in-the-loop strategy, we excluded the 15% of predictions with the highest uncertainty, which increased accuracy for the remaining predictions to 85% (<i>p</i> = 0.006). Model inspection identified Barthel Index, Clinical Frailty Scale, age, and number of medications as key predictors.</p> Conclusion <p>Our findings provide a proof-of-concept that an ML-based approach can effectively support individualized discharge planning for geriatric surgical patients. With 82% congruence with expert recommendations, the model significantly outperformed the standard of care. Its real-world clinical impact is currently being evaluated in an ongoing interventional trial.</p> Trial registration <p>German clinical trials registry (DRKS00030684), registered on 21st November 2022.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Standard-of-care vs. machine learning-recommended discharge destinations for geriatric surgical inpatients: algorithm development and internal validation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Thomas Derya Kocar,
  • Utz Lovis Rieger,
  • Filippo Maria Verri,
  • Pauline Maier,
  • Christoph Leinert,
  • Simone Brefka,
  • Christian Bolenz,
  • Nuh Rahbari,
  • Florian Gebhard,
  • Dhayana Dallmeier,
  • Michael Denkinger,
  • Hans Kestler

摘要

Purpose

Older adults undergoing surgery are at high risk for adverse outcomes due to age-related vulnerabilities. Comprehensive geriatric assessment (CGA) can improve discharge planning but is often limited by resource constraints. We evaluated whether machine learning (ML) can support individualized discharge destination recommendations for geriatric surgical inpatients.

Methods

We developed and internally validated an AdaBoost classifier using electronic health record and CGA data from 169 patients aged ≥ 70 years undergoing surgery at a single center. The model predicted discharge destinations (home, acute geriatric care unit, rehabilitation facility, nursing home) and was evaluated using accuracy, receiver operating characteristic (ROC) curves, and calibration within a fivefold cross-validation framework. Model performance was also compared with previously reported standard-of-care discharge decisions.

Results

The study population consisted of 169 older adults, with mean age of 80.5 years (SD ± 6.3). The ML model achieved a micro-averaged ROC area under the curve of 0.94 [95% CI 0.92–0.96] and an accuracy of 82% [95% CI 0.76–0.86], representing a significant improvement over standard-of-care decisions (138 vs 124 correct, p = 0.041). Mimicking a human-in-the-loop strategy, we excluded the 15% of predictions with the highest uncertainty, which increased accuracy for the remaining predictions to 85% (p = 0.006). Model inspection identified Barthel Index, Clinical Frailty Scale, age, and number of medications as key predictors.

Conclusion

Our findings provide a proof-of-concept that an ML-based approach can effectively support individualized discharge planning for geriatric surgical patients. With 82% congruence with expert recommendations, the model significantly outperformed the standard of care. Its real-world clinical impact is currently being evaluated in an ongoing interventional trial.

Trial registration

German clinical trials registry (DRKS00030684), registered on 21st November 2022.