JMIR Publications

ABSTRACT

Background:

There is tremendous value in being able to predict post-operative complications before and during surgery. This may allow for prevention of morbidity or mortality by the healthcare team and ultimately improve patient outcomes. While there are multiple scoring systems available to aid the physician in determining the probability of post-operative complications, none of the currently available models focus on the individual patient’s likelihood of requiring lifesaving interventions such as continued ventilatory support, extracorporeal membrane oxygenation, or continuous renal replacement therapy in the post-operative setting.

Objective:

Our objective is to develop clinical models that can predict whether a patient will require the above-mentioned life-saving interventions in the postoperative setting through the use of machine learning. This allows for highly individualized risk prediction, allowing a care provider to prevent, prepare, and rapidly address life-threatening postoperative complications.

Methods:

Four artificial neural networks (ANN) and random forest models are utilized to predict the necessity of each of the four life-saving medical devices described above. Preoperative lab data, medication data, and surgery type served as the input to the model, along with intraoperative vital signs. Each network was trained using a dataset named INSPIRE (Lim et. al), which is composed of data from 130,000 patients undergoing surgery from various surgical specialties. Model analysis wass done through both accuracy and area-under-the-curve analysis.

Results:

Each model demonstrated very high efficacy: Extracorporeal membrane oxygenation prediction (Accuracy: 98.9% (ANN) & 98% (Random Forest classification), AUC: .992); Ventilatory support prediction (Accuracy: 97.69% (ANN) & 99% (Random Forest classification), AUC: 0.995); Intra-aortic balloon pump (Accuracy: 98.03% (ANN) & 98.2% (Random Forest classification), AUC: 0.9921); Continuous renal replacement therapy (Accuracy: 94.93% (ANN) & 96.45% (Random Forest Classification), AUC: 0.9833). Evaluation of the random forest decision tree indicates the need of individualized patient risk prediction, an aspect of risk prediction not seen in prior art.

Conclusions:

Utilization of this system can lead to highly accurate post-surgical risk prediction, and can allow for more robust post-operative patient management- thus potentially decreasing morbidity and mortality in patients undergoing both elective and non-elective surgery.