Developing an SSI risk score incorporating daily objective wound assessments using machine learning
Author(s):
Patrick Sanger, University of Washington Department of Biomedical Informatics; Gabrielle van Ramshorst, Department of Surgery, VU University Medical Center Amsterdam, Netherlands; Ezgi Mercan, University of Washington Department of Computer Science and Engineering; Meliha Yetisgen, University of Washington Department of Biomedical Informatics; Andrea Hartzler, Group Health Research Institute; Cheryl Armstrong, University of Washington Department of Surgery; Ross Lordon, University of Washington Department of Biomedical Informatics; Sarah Han, University of Washington Department of Surgery; William Lober, University of Washington Department of Biobehavioral Nursing & Health Systems; Heather Evans, University of Washington Department of Surgery
Background: Surgical site infection (SSI) remains a common, costly and morbid healthcare-associated infection. Prediction of SSI may facilitate earlier recognition and treatment, yet previous SSI risk scoring systems only consider baseline risk factors (BF) on the day of operation, not accounting for changing risk over time after surgery.
Hypothesis: Incorporation of daily wound assessment improves the accuracy and timeliness of SSI prediction compared to traditional BF alone.
Methods: A prospective cohort of 1000 patients scheduled for open abdominal surgery at an academic teaching hospital were examined daily until discharge. Patients who didn’t undergo surgery (n=33) or with <2 days of wound observations (n=107) were excluded. We collected patient and procedure BF and compared SSI vs. non-SSI groups using univariate methods. Daily observations of wound features (e.g., exudate) were recorded. Our primary outcome was CDC-defined SSI. Using supervised machine learning, we trained 3 Naïve Bayes classifiers incorporating correlation-based feature subset selection, evaluated using 10-fold cross validation: one with BF, one with repeated features (RF) and one with both. To train the classifiers, patient data from 1-3 days prior to SSI were used to predict diagnosis on day 0. For patients without in-hospital SSI, we matched 3 similar consecutive post-op days. Accuracy, predictive values and AUC were calculated.
Results: Of 860 patients included in analysis, 20.3% had in-hospital SSI. Mean prediction day for patients who developed SSI vs. no SSI: 7.25 vs. 7.29.
Univariate analysis of SSI vs. non-SSI groups showed differences in c-reactive protein (5.3±8.4 vs 2.6±5.9 mg/dL, p<0.05), surgery duration (308±148 vs. 253±124 min, p<0.05) and contamination (dirty cases 21[11.8%] vs. 25[3.9%], p<0.001), but no differences in ASA scores, diabetes or emergency surgery.
Table 1
Dataset
|
Accuracy
|
Kappa
|
Sens
|
Spec
|
PPV
|
NPV
|
AUC
|
BF
|
77%
|
0.13
|
18%
|
92%
|
38%
|
82%
|
0.67
|
RF
|
78%
|
0.34
|
51%
|
85%
|
46%
|
87%
|
0.76
|
BF+RF
|
79%
|
0.36
|
51%
|
86%
|
48%
|
87%
|
0.76
|
Classifier performance is compared in Table 1. In order, the most predictive features in the RF classifier were granulation score, pulse rate, presence of NG tube, amount of exudate, wound culture ordered, and wound separation distance.
Conclusions: Repeated features provided moderate PPV and high NPV for prediction of SSI in advance of clinical diagnosis. Addition of baseline patient/operative data did not improve prediction. Features of evolving wound infection are discernable prior to the day of diagnosis primarily based on visual inspection.