Despite the advances in peri- and intraoperative prevention measures, surgical site infections (SSIs) remain one of the most common adverse events that occur in patients undergoing in-patient or outpatient surgical procedures. SSIs can be classified as either superficial-incisional (involving the skin or subcutaneous tissue layers of the incision), deep-incisional (involving muscle or connective tissue layers of the incision), and organs/spaces deep to the incision that were opened or manipulated during surgery (Table 1) (1). Of the three, superficial incisional SSI are more common than deep incisional and organ/space SSI (2, 3), and account for more than half of all SSIs for all surgery categories (e.g., cardio-thoracic, gynecologic, orthopedics, transplant) (2).

Regardless of SSI classification, SSIs place a significant burden on both the patient and health system. In the US alone, SSI extends the hospital length of stay by 9.7 days (4) and is associated with a 2- to 11- fold increase in the risk of mortality (5, 6). SSI complications translate into $10 billion in additional healthcare costs (5, 7), with an average of $25,000 excessive cost per case (5, 7).

Focusing specifically on antibiotics to treat SSI, Jenney et al. (8) reported that the wholesale cost of antibiotics to treat SSI after coronary artery bypass surgeries (CABGs) was AUD $391 (AUD $727.43 inflated to 2022 prices). The cost of antibiotic treatment for patients undergoing vascular surgery who developed an SSI was £3,776 (£4,280.25 inflated to 2023 prices) (9). Antibiotics after major head and neck surgery due to MRSA infection was £260 (£429.54 inflated to 2023 prices) on first admission and £1,700 (£2,808.56 inflated to 2023 prices) on each re-admission (10). As such, SSI imposes an enormous clinical and economic burden on both patients and healthcare systems.

Over the past two decades, healthcare systems have implemented strategies that minimize the length of hospital stay and move inpatient surgical procedures to the outpatient setting. Consequently, a growing proportion of SSIs are detected only after discharge (11). Indeed, of the 500,000 SSIs that occur in the U.S. annually, an estimated 69% of SSIs occur after hospital discharge (11). At present, there is no international scientific consensus about the optimal method for post-discharge SSI surveillance (12). Common methods for identifying surgical wound infection after hospital discharge include direct observation by healthcare professional, telephone interview, patient questionnaire, and outpatient clinic follow-up (13). Because post-discharge surveillance remains unstructured, SSIs are often overlooked (14, 15), so there may be a significant time delay before the physician is able to detect the infection. This is unfortunate, given that a 45-hour delay in detection and treatment of an SSI increases the odds of infection related deaths by 3.8 times (6, 16, 17).

With increasing demand for healthcare, SSIs need to be predicted and diagnosed early so that timely and effective treatment (e.g., antibiotics) can be implemented to accelerate the recovery of patients (4, 18–20). Fortunately, normal surgical wound healing processes occur in a defined and organized fashion, requiring oxygen to promote constructive healing and fight off any contaminants or organisms (21). The development and/or presence of infection leads to directly observable changes (e.g., swelling and erythema) (22) that directly reflect changes in tissue physiology (e.g., abnormal tissue oxygenation, altered acid base, and temperature changes) (23). Such changes can be measured using sensor-enabled technologies with sophisticated signal processing techniques (24–28). For example, Govinda et al. [2010] reported that upper arm subcutaneous oxygen partial pressure recorded using near-infrared spectroscopy 75 min after colorectal surgery predicted SSI with a sensitivity of 71% and specificity of 60%. More recently, Mostafalu et al. [2018] developed a wound dressing capable of continuously measuring wound pH and temperature connected to a wireless electronic module, with in vitro bacterial testing indicating that the system could accurately and reliably record wound potential of hydrogen [pH] and temperature.

Taken together, these contribute to the growing evidence that the status of wound healing can be objectively quantified using digital devices. The current study builds on this research by developing a multi-modal bio-signal system equipped with clinical-grade sensors capable of continuously monitoring peri-wound tissue oxygen saturation (StO2), temperature, and bioimpedance (BioZ). The ability of the developed system to predict developing infection in a porcine model was evaluated using a bagged, stacked, and balanced ensemble logistic regression machine-learning model. These results are the first step towards determining the efficacy of a multi-modal sensor system to collect digital biomarkers continuously from the surgical site to monitor the healing status of the wound, with the goal of eventually developing a machine-learning based early warning system to detect and predict SSIs that could become the standard of care for remotely monitoring patients' post-surgery. The aim of the present study was to evaluate the ability of a multi-modal bio-signal acquisition system to predict current and developing superficial incisional infection in a porcine model infected with Methicillin Susceptible Staphylococcus Aureus (methicillinsusceptible S. aureus, MSSA).

A total of 14 incisions were made on each pig: seven were inoculated with MSSA, six served as sham sites, and one served as a control site. For pig #1, 5/7 (71.4%) inoculated sites were infected, 1/6 sham wound sites were infected (14.3%), and 0/1 (0%) control site were infected. For pig #2, 6/7 (85.7%) inoculated sites were infected, 0/6 sham wound sites were infected (14.3%), and 0/0 (0%) control sites was infected.

A total of 28 Hematoxylin and Eosin (H&E) stained skin tissue samples were scored by a certified Research Veterinary Pathologist for histopathological changes such as inflammation, granulation tissue, collagen deposition, epithelization and neovascularization based on the score pattern described by Barington et al. (40). The infiltration of polymorphonuclear cells (PMNCs) persist in inoculated biopsy skin samples from days one though six, decreasing in distribution until day five, with a sudden increase in presence on days three and six. The infiltration of PMNCs in sham sites is decreased compared to inoculation sites on various days of surgical removal of biopsy. The presence of mononuclear cells in inoculation sites increased from day 1–5, and comparatively declined in distribution in sham sites. Samples taken from the control site failed to show pathological changes.

Figure 3 (top panel) illustrates time-related differences in peri-wound tissue oxygen saturation (% StO2), wound temperature (temperature from baseline), and bioimpedance (BioZ from baseline) for non-infected (green) vs. superficial incisional infected wounds (red) relative to the sham wounds. Statistical differences between wounds are indicated with black rings (i.e., ○). The percentage of tissue oxygenation (% StO2) for infected wounds followed a bell curve pattern under 72 h post-inoculation peaking at the 25th hour. This differed from non-infected wounds that were relatively stable across the study period. Statistical analysis indicated that differences in percent StO2 between the infected and non-infected surgical sites reached statistical significance at 12 h post-inoculation until 50 h post-inoculation (Figure 3 bottom left panel, all p's < 0.05). Wound temperature (°C from baseline) between non-infected and infected wounds demonstrated similar trends over the study period, but with markedly different values that started 18 h post-inoculation and lasted until 96 h post-inoculation (Figure 3 bottom middle panel). However, statistical analysis indicated that the differences in wound temperature between the infected and non-infected surgical sites reached statistical significance only during the time period between 24 until 72 h post-inoculation (all p's < 0.05). Wound bioimpedance (Ohm from baseline) between non-infected and infected wounds demonstrated similar trends over the study period, but with markedly different values that started 48 h post-inoculation and lasted until 120 h post-inoculation (Figure 3 bottom right panel). However, statistical analysis indicated that the differences in wound bioimpedance between the infected and non-infected surgical sites reached statistical significance only during the time period between 72 until 120 h post-inoculation (all p's < 0.05).

For superficial incisional infected wounds, cross-correlational analysis indicated a change in the three bio-signals that occurred at least 24 h before a change in veterinarian-based SSI diagnosis (Figure 4). The percentage of tissue oxygenation (% StO2) for infected wounds exhibited a significant upward trend 25 h in advance of a clinician's diagnosis of wound infection (r = 0.39, p < 0.001). Similarly, the peri-wound temperature exhibited a moderately upward trend at 31 h before a diagnosis of infection (r = 0.40, p < 0.001). A moderately strong downward trend in the BioZ value of the tissue was observed 24 h prior to the clinical diagnosis of wound infection (r = 0.54, p < 0.001).

Table 3 presents the ML model's ability to predict current and future superficial incisional SSI. The ability of the model to detect a current SSI achieved an AUC of 0.77, accuracy of 64%, sensitivity of 82%, and specificity of 59%. In comparison, the predictive performance of the model was higher when predicting SSI 24 h in advance of veterinarian-based SSI diagnosis, with an AUC of 0.80, accuracy of 78%, sensitivity of 81%, and specificity of 76%. The algorithm exhibited lower performance for predicting SSI 48 h in advance of veterinarian-based SSI diagnosis (AUC = 0.74, accuracy = 69%, sensitivity = 77%, specificity = 65%). By pairing and comparing the AUC using DeLong's test, we found that there was no statistical difference between predicting current SSI and 24 h in advance of veterinarian-based SSI diagnosis (p = 0.071), between current and 48 h in advance of veterinarian-based SSI diagnosis (p = 0.0790), while the difference between predicting SSI 24 h in advance of veterinarian-based diagnosis and predicting SSI 48 h in advance of veterinarian-based SSI diagnosis was statistically significant (p < 0.001).

The aim of the present study was to evaluate the ability of a multi-modal bio-signal acquisition system to predict current and developing superficial incisional infection in a porcine model infected with Methicillin Susceptible Staphylococcus Aureus. The predictions of the system were evaluated using a bagged, stacked, and balanced ensemble logistic regression machine learning model.

The acquired bio-signals enabled the characterization of wound healing and superficial incisional infection from MSSA across the study period. At non-infected wound sites, StO2 was relatively stable from the point of inoculation to the end of the study period, temperature increased from the point of inoculation until the 24th hour then decreased gradually, and bioimpedance decreased from the point of inoculation to the 80th hour, then increased to the end of the study period. In contrast, StO2 of infected wounds increased from the point of inoculation to the 25th hour then decreased to the 80th hour. Temperature and bioimpedance of infected wounds showed similar trends to the non-infected wounds, but with higher peri-wound temperatures from the 24th hour through the 96th hour, and lower bioimpedance values from the 48th hour through the 120th hour. When comparing the values of the individual biomarkers to clinical wound scores, the results of the present study demonstrate that digital signals can indicate the presence of an MSSA infection 24 to 31 h before standard methods. The ability to detect changes in tissue physiology associated with SSI is of critical importance to healthcare providers, as it would help enable the timely and effective treatment (e.g., antibiotics) to accelerate the recovery of patients (4, 20), that would ultimately reduce healthcare costs, SSI-associated complications and mortality rates.

Interestingly, there were statistically significant differences between the inoculated and sham wound sites 60–55, 10, and 2–1 h prior to the incision and introduction of MSSA. Because the distribution of inoculated and sham wounds across the trunk of the animals, it is unlikely to be due to local physiological effects. Rather, because the p-values during these time periods range from p = 0.041 to 0.047 we believe that the statistical significance at these time periods are spurious in nature, and be non-significant with more data. We minimized the number of animals involved in the current study, because the experimental research methodology was invasive by its very nature which resulted in potential undue distress and harm on the animals. As the research moves into post-surgical human populations, the non-invasive observational research methodologies (i.e., monitoring post-surgical patients) will allow us to evaluate the system in a greater number of wound sites. If our assumption regarding statistical spuriousness is correct, then we expect that pre-surgical temperature will be similar within an anatomical region.

Congruent with prior studies that have developed machine-learning powered SSI models using medical data (41) or wound imaging (42), the multi-modal ensemble model developed in this study demonstrated promising accuracy and sensitivity to detect a current infection, yielding an AUC value of 0.77. While model performance to detect SSI 24 h and 48 h in advance of veterinarian-based SSI diagnosis (AUC's = 0.80 and 0.74, respectively) also demonstrated acceptable discriminability, DeLong's tests indicated that both the current infection and the 24 h in advance models performed equally well, whereas the model predicting SSI 48 h in advance of veterinarian-based SSI diagnosis performed significantly worse. Based on the data, it can be put forth that the measured bio-signals are best at providing information regarding pathophysiological processes before overt clinical signs are apparent. This conjecture is supported by results of the cross-correlation analysis indicating that peri-wound bioimpedance and tissue oxygenation of inoculated wounds peaked 25 and 24 h before veterinarian-based SSI diagnosis. Taken together, the current data indicate that superficial incisional wounds infected with MSSA result in physiological changes to the peri-wound area (e.g., rise in peri-wound temperature due to inflammation, immune responses, and/or tissue metabolism) that can be detected using non-invasive multi-modal sensors.

While the model showed acceptable performance, the diagnostic accuracy of the model could be further improved by increasing the volume of training data and/or handling the problem of imbalanced data by using rule-based methods that can learn high confidence rules for the minority SSI class (43) or by building cost-sensitive classifiers (44). Another approach, which may lead to greater clinical adoption, would be to integrate sociodemographic and clinical information (e.g., hypertension, diabetes mellitus, current or past smoker), peri-operative (e.g., urgency of surgery, duration of surgery, surgical drain placed, antibiotic prophylaxis, intraoperative supplemental oxygen), and post-operative parameters (e.g., serous exudate, purulent exudate, white blood cell count, antibiotics administered) into the model.

There are limitations to the current study that may inform future directions in this line of research. First, the sensors were sutured on the skin of the porcine subjects in order to minimize movement-related measurement errors and signal noise. We have adapted the CrelySENSE sensors for use in clinical trials involving human volunteers so that they attach to the peri-wound site via disposable adhesives and are connected to the CrelyPRO using Bluetooth low energy (BLE) protocols. Second, SSI may also lead to changes in bio-signals such as perfusion, pulse rate, respiration rate, and pH. The inclusion of these parameters may improve the detection and prediction of SSIs in animal and human populations. Third, we examined responses after MSSA infection. While the most common organism that causes SSIs (45–47), it is unknown whether SSIs caused by bacteria such as Methicillin-resistant Staphylococcus aureus (MRSA), extended-spectrum β-lactamase (ESBL)-producing Enterobacteriaceae [Escherichia coli (E. coli), Klebsiella pneumoniae], multidrug resistant gram-negative bacteria (e.g., Acinetobacter baumannii), and vancomycin-resistant Enterococcus faecium (VRE) would yield similar bio-signal responses. Lastly, the current study focused on the detection and prediction of superficial incisional SSIs that occur in the area of the skin where the incision was made. Whether or not the described bio-signal system can detect infections that occur beneath the incisional area in the muscle and surrounding tissues (deep incisional SSIs) remains to be ascertained. There is some evidence that post-surgical deep incisional infections may exhibit different patterns of expression of wound biomarkers (48, 49). Future studies should thus validate the current system in different types of SSIs and pathogens.

In summary, the results of the current study bear out the accuracy, sensitivity, and specificity with which the non-invasive multi-modal bio-signal system of interest is capable of detecting MSSA infection in porcine subjects. Such systems can provide clinicians with a continuous and real-time understanding of the state of the wound healing process, enabling them to make timely decisions regarding wound care that would substantially reduce post-operative complications, length of stay, and treatment costs.