Optimal computed tomography-based biomarkers for prediction of incisional hernia formation

Abstract

Purpose

Unstructured data are an untapped source for surgical prediction. Modern image analysis and machine learning (ML) can harness unstructured data in medical imaging. Incisional hernia (IH) is a pervasive surgical disease, well-suited for prediction using image analysis. Our objective was to identify optimal biomarkers (OBMs) from preoperative abdominopelvic computed tomography (CT) imaging which are most predictive of IH development.

Methods

Two hundred and twelve rigorously matched colorectal surgery patients at our institution were included. Preoperative abdominopelvic CT scans were segmented to derive linear, volumetric, intensity-based, and textural features. These features were analyzed to find a small subset of OBMs, which are maximally predictive of IH. Three ML classifiers (Ensemble Boosting, Random Forest, SVM) trained on these OBMs were used for prediction of IH.

Results

Altogether, 279 features were extracted from each CT scan. The most predictive OBMs found were: (1) abdominopelvic visceral adipose tissue (VAT) volume, normalized for height; (2) abdominopelvic skeletal muscle tissue volume, normalized for height; and (3) pelvic VAT volume to pelvic outer aspect of body wall skeletal musculature (OAM) volume ratio. Among ML prediction models, Ensemble Boosting produced the best performance with an AUC of 0.85, accuracy of 0.83, sensitivity of 0.86, and specificity of 0.81.

Conclusion

These OBMs suggest increased intra-abdominopelvic volume/pressure as the salient pathophysiologic driver and likely mechanism for IH formation. ML models using these OBMs are highly predictive for IH development. The next generation of surgical prediction will maximize the utility of unstructured data using advanced image analysis and ML.

Introduction

Surgical complications are a major burden on patients and the health care system [1]. Generally speaking, current surgical prediction techniques are reliant on organized and easily interpretable information stored within Electronic Health Records (EHRs) [2–4], which we will refer to as structured data. However, these data are often limited in scope and may not be appropriate surrogates for underlying drivers of surgical outcomes. There are potentially revealing factors in unstructured data (e.g., medical images and clinical notes). In fact, it is estimated that 80% of healthcare data are of this latter form [5]. With the power of modern analysis techniques, we can now harness information from unstructured data. In particular, advanced image analysis and machine learning (ML) techniques allow us to systematically decipher medical images to improve understanding of surgical disease with potential implications for clinical practice [6–9]. Doing so can elucidate key proximate drivers for surgical outcomes related to a patient’s body habitus and improve surgical prediction.

Incisional hernia (IH) is a quintessential surgical complication amenable to exploration of unstructured data, as it is a common, long-term complication of abdominopelvic surgery that is incompletely understood [10–12]. IH is hallmarked by failed fascial healing resulting in protrusion of the abdominopelvic contents, leading to chronic pain, diminished physical function, poor mental health, and unsuitable appearance [13–17]. IH is a prevalent public health concern because it is estimated 1-in-8 abdominal surgery patients develop it [10, 12, 18–20]. Furthermore, nearly one-third of IH repairs fail, leading to a vicious cycle of disability and costs [15, 21, 22]. For surgeons to employ risk mitigation strategies, prediction models, such as the Penn Hernia Risk Calculator, have been developed using hundreds of structured data elements, yet have modest prediction ability [4, 23–25]. Features from preoperative medical imaging may elucidate proximate drivers of IH, enable deeper understand of its pathophysiology, and, thus, better prediction and prevention efforts.

We have developed methods to systematically derive image-based features from preoperative abdominopelvic computed tomography (CT) imaging of abdominal surgery patients [6, 7, 26]. Our recent pilot study identified some features which have discriminative capacity for IH [26]. The primary objective of this study is to adapt the optimal biomarker (OBM) method [8, 9] to identify key image-based features from preoperative abdominopelvic CT imaging with the best risk prediction abilities for IH. The second objective was to employ various ML classifiers to predict IH formation using these OBMs and compare performance.

Methods

Patient selection and image analysis

This study was approved by the Institutional Review Board (IRB) at the University of Pennsylvania (Protocol # 843,902). Informed consent was waived by the IRB due to the retrospective and deidentified nature of the study. Two hundred and twelve adult patients undergoing elective colorectal surgery at our institution between 2005 and 2016 were matched 1:1 for the outcome of developing IH. Patients were exactly matched on 10 preoperative characteristics important to IH risk: age (± 4 years), sex, race, ethnicity, smoking status, history of diabetes mellitus, pre-index ventral incisional hernia repair, pre-index abdominal surgery, index laparoscopic surgery, and index open surgery. Their preoperative CT scans were retrospectively obtained and evaluated for 107 linear and morphometric features as in our previous work [26]. The same CT images were now assessed additionally for textural and intensity-based features. These tissual properties were evaluated as mean, median, standard deviation, first quartile, third quartile, maximum, and minimum for various tissues in the abdomen, pelvis, and abdomen-pelvis. In aggregate, there were 84 texture-based features and 42 intensity-based features.

Proportional measurements between morphometric features were also considered as additional derived features, such as visceral adipose tissue (VAT) volume as a fraction of inner aspect of body wall skeletal musculature (IAM) volume within the abdomen, pelvis, and abdomen-pelvis. Examples of delineated regions on slices for which volumetric measurements are computed are depicted in Fig. 1 (adapted with permission from McAuliffe, et al.) [26]. In total, there were 44 ratio features considered. All previous and new features were included in the analysis which resulted in a total of 279 features.

Fig. 1.

CT-derived volumetric measurements of patient abdominopelvic body habitus. Delineations overlaid on CT slice displays show tissue regions from which volumetric measurement are derived. IAM inner aspect of the body wall skeletal musculature, OAM outer aspect of the body wall skeletal musculature, Skin skin and interior contents, VAT visceral adipose tissue, SMT skeletal muscle tissue, SAT subcutaneous adipose tissue. Adapted with permission from McAuliffe, et al. [26]

Optimal biomarker (OBM) methodology

A general Optimal Biomarker (OBM) methodology was previously developed by our team and validated in different applications to identify an optimal set of predictive features [8, 9, 27]. The OBM methodology finds a small subset from a large set of features that are both discriminative for the outcome of interest and independent from other features. In the current study, this meant finding features which (i) had a low magnitude of correlation (r ≤ 0.5) with most of the other features in the given set; (ii) were maximally discriminative in separating patients who developed IH from those who did not, as determined by p value; and (iii) which provided the best prediction accuracy when used with all other features of the subset in a ML prediction model. While there were 279 features identified per CT, there are no known practical methods to find the best subset among 2²⁷⁹− 1 possible subsets. Therefore, we instead identified the subset of the most potent six features as an approximate solution [8]. In addition, in this work, we modified the original OBM method to handle data given as matched pairs. The details of the implementation are as follows.

Our matched cohort of 212 patients (106 pairs) was randomly partitioned into a 128:42:42 training:validating:testing split wherein pairing was maintained, with the ratio for pairs being 64:21:21. The training set was used to determine the top six most potent among 279 features following the aforementioned OBM methodology such that the features are (i) least correlated and (ii) most discriminative.

Once the top six most potent features were identified, an ensemble boosting ML classifier was then trained on the training data set separately for each subset among the 2⁶− 1 = 63 possible subsets of the six-feature set. The resulting 63 models were tested on the validation data set for predicting IH formation. The subset of features with the highest prediction accuracy on the validation data set was considered to represent the OBMs.

Finally, three ML classifiers—support vector machine (SVM), random forest, and ensemble boosting—were trained on the training + validation data sets (170 scans or 85 pairs) for the OBMs found for final predictive performance evaluation on the remaining 42 scans (21 pairs) set aside solely for testing [28–30].

To handle paired data effectively, the feature values were transformed from the measured values to the difference between the measured value and the mean of its pair. In mathematical terms, for each feature f among the 279 features, let its paired values be (a, b), where a and b denote the values of f for the Hernia and no-Hernia groups, respectively. (a, b) is modified to the pair (a’, b’) where a’ = a − (a + b)/2 and b’ = b − (a + b)/2. In other words, a’ and b’ denote the signed distance of a and b, respectively, from the mean. These transformed values were used for training, validation, and testing experiments.

Results

The OBM method revealed that the most potent CT features for identifying IH are all morphometric in nature: abdominopelvic VAT volume normalized for abdominopelvic height, abdominopelvic IAM volume normalized for abdominopelvic height, pelvic VAT volume normalized for pelvic height, abdominopelvic IAM/OAM volume ratio, pelvic VAT/OAM volume ratio, and pelvic IAM volume normalized for pelvic height. Table 1 lists these features and their summary statistics for the two groups.

Table 1.

Top six most potent features and their mean and standard deviation for the two classes for prediction of incisional hernia

Top six potent feature	No hernia [mean (SD)]	Hernia [mean (SD)]	p-value
AbdPlvs VAT vol normalized (m3)	0.06 (0.04)	0.08 (0.04)	0.02
AbdPlvs IAM vol normalized (m3)	0.18 (0.05)	0.19 (0.05)	0.03
Plvs VAT vol normalized (m3)	0.11 (0.06)	0.13 (0.07)	0.01
AbdPlvs IAM/OAM	0.46 (0.05)	0.48 (0.06)	0.02
Plvs VAT/OAM	0.12 (0.06)	0.14 (0.06)	0.02
Plvs IAM vol normalized (m3)	0.29 (0.08)	0.32 (0.08)	0.01

Abd Abdomen, Plvs pelvis, AbdPlvs abdomen-pelvis, VAT visceral adipose tissue, IAM inner aspect of body wall skeletal musculature, OAM outer aspect of body wall skeletal musculature, Vol volume, Normalized normalized to body region height, SD standard deviation

The six most potent feature biomarkers are all related to three fundamental volumes of the lower abdomen and pelvis: VAT, IAM, and OAM volumes. Figure 2 shows three-dimensional depictions of the surfaces of these volumes from CT images in two matched-pair of subjects, one of whom developed IH and the other who did not. Videos 1 and 2 are complete views of these three-dimensional renditions of key body volumes. The validation stage for the best ensemble boosting classifier is depicted in Table 2, where the best subset found with 1, 2, …, 6 features is shown along with the corresponding accuracy, sensitivity, specificity, and AUC metrics. The best subset included three features (i.e., OBMs): abdominopelvic VAT volume normalized for abdominopelvic height, abdominopelvic IAM volume normalized for abdominopelvic height, and pelvic VAT/OAM volume ratio. These OBMs from the training and validation sets were used to train the three ML classifiers, with ROC curves for prediction on the testing set shown in Fig. 3. The accuracy, sensitivity, specificity, and AUC of the classifiers are listed in Table 3. The ensemble boosting classifier yielded the best performance across all prediction metrics for the paired approach.

Fig. 2.

Optimal volumetric features for prediction of incisional hernia (IH), depicted from a sample of matched patients as 3D rendered images of the surfaces of the volumetric objects. VAT visceral adipose tissue, OAM outer aspect of the body wall skeletal musculature, IAM inner aspect of the body wall skeletal musculature

Table 2.

Performance of best combinations of top features using an ensemble classifier on the validation cohort for prediction of incisional hernia

The number of features in best subsets	Index of features in best subsets	Accuracy	Sensitivity	Specificity	AUC
1	19	0.69	0.62	0.76	0.79
2	19, 107	0.81	0.81	0.81	0.85
3	109, 19, 120	0.83	0.86	0.81	0.85
4	109, 19, 107, 120	0.83	0.86	0.81	0.84
5	109, 19, 107, 112, 120	0.81	0.81	0.81	0.82
6	109, 19, 107, 112, 120, 1 3	0.67	0.71	0.62	0.73

The optimal biomarkers are those in the best performing subset of three, as this subset has the best accuracy, sensitivity, specificity, and AUC (bolded)

Feature 19–AbdPelvis IAM volume norm, Feature 107–Pelvis VAT volume norm, Feature 109–AbdPelvis VAT volume norm, Feature 120–Pelvis VAT/OAM volume ratio, Feature 112–AbdPelvis IAM/OAM volume ratio, Feature 13–Pelvis IAM volume norm

Abd Abdomen, Plvs pelvis, AbdPlvs abdomen-pelvis, VAT visceral adipose tissue, IAM inner aspect of body wall skeletal musculature, OAM outer aspect of body wall skeletal musculature, Vol volume, Norm normalized to body region height, AUC area under the curve

Fig. 3.

Receiver operating characteristic (ROC) curves of three machine learning models trained using the best combination of optimal biomarkers for prediction of incisional hernia. SVM support vector machine, AUC area under the curve

Table 3.

Prediction of incisional hernia using combined three optimal biomarkers, using three machine learning classifiers: ensemble boosting, random forest, SVM

Classifier	Accuracy	Sensitivity	Specificity	AUC
Ensemble boosting	0.83	0.86	0.81	0.85
Random forest	0.79	0.76	0.81	0.83
SVM	0.67	0.67	0.67	0.68

The ensemble boosting classifier model had the best performance in terms of accuracy, sensitivity, specificity, and AUC (bolded)

SVM Support vector machine, AUC area under the curve

Discussion

This study demonstrates that morphometric image-based features may denote key drivers of surgical risk and IH formation. Moreover, these optimal features characterize a singular physiologic domain, increased intra-abdominopelvic volume, which likely represents a more proximate pathophysiologic driver of IH. While results are preliminary, the study finds that ML modeling of IH using these optimal image-based features is not only possible, but also can predict IH formation with excellent accuracy, albeit in a limited cohort. The results further suggest that image-based prediction pairing OBMs with ML modeling are crucial in augmenting our understanding and management of surgical outcomes. Such features must be validated in a larger cohort and can eventually be integrated to build a comprehensive risk model using healthcare datasets of both structured and imaging data elements.

Our previous work determined that biometric features discriminative for IH formation fell into three pathophysiologic domains: widening of the rectus complex and linea alba, increased visceral volume, and atrophy of the abdominopelvic wall skeletal musculature [26]. The current study significantly expands upon this pilot work by finding the six most optimal biomarkers from CT image analysis. All six features characterize a single physiologic domain: increased intra-abdominopelvic volume. This suggests that increased intra-abdominopelvic volume is a critical pathophysiologic mechanism for IH formation. Previous literature agrees with this notion as it has been suggested that visceral obesity is a mechanistic driver of IH formation [31–33].

Further investigation must be done to mechanistically connect increased visceral volume with IH. We propose two potential explanations. The first is that increased visceral adiposity may contribute to IH directly through molecular mechanisms of wound healing. Abdominal adipocyte hypertrophy typical of obesity is coupled with inadequate vascularization [34]. The resulting hypoxia can delay wound healing and lead to a pro-fibrotic environment. This dovetails well with existing literature establishing impaired wound healing as mechanistically significant in IH pathogenesis. Altered collagen subtypes and fibroblast phenotypes, resulting in altered expression of matrix metalloproteinases (MMPs) and tissue inhibitor of metalloproteinases (TIMPs), lead to disorganization of the wound extracellular matrix [35–37]. This results in impaired fascial healing and decreased integrity of the abdominal wall. Another possible explanation is that the optimal biomarkers indicating increased intra-abdominopelvic volume may actually be a proxy for increased pressure, possibly an even more proximate driver of IH. Such a mechanical stimulus may affect fibroblast homeostasis during the fibroproliferative phase of wound healing [35]. In an analogous finding, Besancenot et al. recently identified chronic obstructive pulmonary disease (COPD) as a risk factor for IH, again possibly due to intrathoracic air trapping referring increased pressure in the abdomen [38]. This increased intra-abdominopelvic pressure may also cause early micro-gapping and failed fascial healing. More exploration is needed to better elucidate the mechanistic pathways, to identify whether it is intra-abdominopelvic volume or pressure that is the more proximate driving force for IH, and whether there are ways to mitigate this biomechanical tendency peri-operatively to prevent IH.

The current study found ensemble boosting to be the most accurate ML classifier for predicting IH using image-based features. Ensemble methods are a machine learning technique which iterate multiple ML algorithms over datasets, evaluate each base model, and combine the base models into a larger ensemble model to perform more effective classification [39–42]. Indeed, the results find that ensemble boosting outperforms other ML methods, such as random forests and SVMs. The best risk model derived in this study using just three image-based features, with an AUC of 0.85, compares favorably with the Penn Hernia Calculator, a model using on hundreds of clinical data points derived from thousands of patients, with an AUC of 0.83 [4]. While the comparison is not entirely fungible given the stringent criteria and preliminary nature of the current study, it demonstrates the potential of harnessing image-based data in surgical risk prediction. This study provides the groundwork for future image analysis to employ ensemble boosting methods with other image-based features.

This study creates high-performing predictive models using exclusively image-based features. However, the results are still very preliminary for ultimate clinical translation. This study uses a small, matched cohort of patients from a single academic institution, which may confound signals of potentially important image-based features. Future work must be done using a larger cohort of surgical patients with preoperative imaging to better establish the OBMs. With greater signal, it is possible that other features related to morphometry, intensity, and texture from CT imaging may become more important in predicting IH. After establishing in a larger cohort, the OBMs must be externally validated across surgical specialties and centers. In this future work, other types of machine learning models, which take advantage of ensemble methods, should be developed using these OBMs to compare their performance in risk prediction [43]. For clinical translation, this line of work must improve IH risk prediction and mitigation ability. To that end, OBMs will need to be routinely analyzed from preoperative imaging. Other sources of unstructured data, such as clinical documentation or genetic profile, should also be analyzed to reveal even more potential features. The ultimate goal is to build an integrated, multi-modal risk model using multidimensional healthcare data sets constructed of structured and unstructured clinical data from the EHR. Eventually, such a model will facilitate development of a point-of-care tool for surgeons to preoperatively assess patient-specific risk of IH formation and guide care recommendations.

Conclusion

Imaging data are a powerful, untapped source for studying surgical outcomes. This study applied advanced image analysis on preoperative CT images to identify optimal biomarkers (OBMs) predictive of incisional hernia (IH) formation. These OBMs suggest increased abdominopelvic volume is a salient pathophysiologic mechanism for IH. Machine learning models trained on these OBMs can provide high-performing risk prediction for IH. The next generation of surgical prediction will maximize the information stored in EHRs in a framework that incorporates both structured and medical imaging healthcare data.

Abbreviations

CT

C omputed tomography

EHR

Electronic health record

IAM

Inner aspect of body wall skeletal musculature

IH

I ncisional hernia

OAM

Outer aspect of body wall skeletal musculature

OBM

O ptimal biomarker

MIPG

Medical image processing group

ML

Machine learning

SVM

Support vector machine

VAT

Visceral adipose tissue

Data availability

Not applicable.