The following is a summary of “Explainable Machine Learning to Bring Database to the Bedside,” published in the July 2023 issue of Surgery by Choi et al.
Illustrate a comprehensible machine learning framework for integrating medical databases into clinical practice; design and verify a point-of-care frailty assessment tool to predict patient outcomes following an injury. An underexplored area in the medical field is the development of a senior trauma frailty index that focuses solely on baseline conditions. This index should be easily implemented and validated on a national scale. Researchers postulated that the Trauma Frailty Outcomes (TROUT) Index has the potential to predict significant adverse outcomes with limited implementation obstacles. They have formulated the TROUT index following the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis guidelines. Utilizing nationwide admission encounters of elderly patients (aged ≥65 years) in the United States from 2016 through 2017, a combination of unsupervised and supervised machine learning algorithms were employed to identify the baseline medical conditions that have the most significant impact on unfavorable outcomes. The dataset was divided into two cohorts: a development cohort comprising 10% of the data and a validation cohort comprising 90%.
These medical conditions were combined to calculate TROUT Index scores (ranging from 0 to 100) that categorize individuals into three levels of frailty risk. Following an examination of the relationship between different levels of vulnerability to frailty and the resulting outcomes, while taking into account factors such as age, gender, and the severity of the injury (as a modifying factor) and conducting calibration analysis, they have developed a mobile application that aims to support the implementation of medical care at the point of care. Investigator’s study population consisted of 1.6 million survey-weighted admission encounters. The TROUT Index consisted of fourteen pre-existing medical conditions and one injury mechanism as the baseline factors. Within the validation cohort, the risk of frailty (with low frailty as the reference group, followed by moderate and high deficiency) was associated with a stepwise increase in the adjusted odds of mortality. The odds ratios (OR) and corresponding 95% CI were as follows: 2.6 (2.4–2.8) for low frailty 4.3 (4.0–4.7) for moderate deficiency. Additionally, an increase in frailty risk was also associated with prolonged hospitalization, with OR (95% CI) of 1.4 (1.4–1.5) for low frailty and 1.8 (1.8–1.9) for moderate weakness.
Furthermore, an increase in frailty risk was linked to a higher likelihood of being transferred to a facility, with OR (95% CI) of 1.49 (1.4–1.5) for low frailty and 1.8 (1.7–1.8) for moderate defect. Lastly, an increase in frailty risk was associated with an increased likelihood of requiring mechanical ventilation, with OR (95% CI) of 2.3 (1.9–2.7) for low frailty and 3.6 (3.0–4.5) for moderate deficiency. The calibration analysis revealed positive correlations between elevated TROUT Index scores and all unfavorable medical outcomes. Study group developed a mobile application called “TROUT Index” and made the source code available to the public. The TROUT Index is a clinically applicable tool used to assess and incorporate frailty into the decision-making process for injured patients. The TROUT Index should not be relied upon as a standalone tool for predicting outcomes following an injury. When utilizing their device, it is essential to consider the institution’s injury pattern, clinical management, and specific workflows. A clinically helpful mobile application and openly accessible code can support future implementation and external validation studies in the medical field.
