The potential of computerised analysis of bowel sounds for diagnosis of gastrointestinal conditions: a systematic review

The aim of this review was to assess the potential for use of bowel sounds computerised analysis in the diagnosis of GI conditions. We looked at the value of specific tests in the clinical setting, as well as the potential of the general approach. Reporting followed PRISMA guidelines. The completed checklist is provided as Additional file 1.

We included studies of any type and in any setting where BSCA was used as a tool for identification or characterisation of GI conditions (clinical trials of devices; diagnostic test accuracy studies, both single gate or case-control; studies with retrospective diagnosis from the data; and preliminary observational studies with tests for associations between bowel sound characteristics and GI conditions or heterogeneity across groups). This reflects the full breadth of studies found along the development pathway of a diagnostic test from proof-of-concept through to widespread clinical use. We included original studies (not reviews) on human subjects, published in a peer-reviewed journal in English. The publication had to have an abstract, to allow the first stage of our search protocol and as an indicator of the depth of the study. We included conference proceedings from the search of engineering journals, given these are typically both comprehensive and peer-reviewed. The participants could be any age, excluding studies with foetal subjects.

Other inclusion criteria were shaped by the PIRD terms for the DTA studies (which we defined as any study that produced accuracy measures, such as sensitivity, specificity) and PICOS terms for the simpler preliminary studies (see Table 1). Our analysis included no limitations on year of publication and inclusion criteria were broad, so as to identify a comprehensive list of relevant studies.

Searches were made of electronic databases with the last search made on 7 April 2017. We also searched for additional studies by reading the references of each included paper.

Two reviewers (AI and KMW) extracted the data independently to standard data extraction forms piloted using three studies. Discrepancies were identified and resolved through discussion. For each study that passed initial screening, reviewed data was collected on the eligibility criteria (see Table 1) and an eligibility decision documented. In addition, data was sought and recorded for the variables in Table 2 for all included studies.

The breadth of studies included made use of a standard quality assessment tool problematic. Hence, bias was assessed at the study level using a modified tool. The QUADAS-2 tool for Quality Assessment for Diagnostic Accuracy Studies was heavily modified (by KMW) with input from three validated checklists used in the quality assessment of a range of study designs for intervention studies: those of Downs and Black [10], Cho and Bero [11], and Moga et al. [12]. The tool was piloted and amended by two independent reviewers (AI and KMW). The resultant tool covered selection bias, performance bias, attrition bias, and reporting bias, as well as competing interests. Across the domains, some questions were retained or added that were applicable to all type of studies, e.g. domain 1: patient selection included ‘Did the study avoid inappropriate exclusions’ (from QUADAS-2) and ‘Were the characteristics of the cohorts clearly described’ (derived from Downs and Black [10]). Consideration of the checklists, especially Cho and Bero [11] and Moga et al. [12], prompted us to add additional questions on statistics and competing interests (see domains 5 and 6).

Elsewhere in the tool, pairs of signalling questions, one question for the DTA studies and one for the preliminary studies testing for associations or differences across groups (see Additional file 2), were used that each addressed the same aspect of bias. For example, the first signalling question for a DTA study was ‘Was a case-control (two-gate) design avoided?’ (derived from QUADAS 2), whilst for a preliminary study, it was replaced with ‘Were the control subjects/cohorts appropriate (similar population to those with the GI condition, matched or random)?’ (derived from Cho and Bero [11]).

Given the simpler preliminary or proof-of-concept studies tend to intrinsically have a higher rate of bias due to a case-control design, the question set used was noted and the results are presented separately below.

Some studies included two components: a test for an association and a subsequent ROC analysis. These were assessed twice, with the appropriate questions for the different parts of the studies.

Quality assessments of all included studies were performed by two independent reviewers (AI and KMW). In the event of disagreement after discussion to reach a consensus, a decision was made by a third reviewer (BJM).

We retrieved studies using the search protocol specified in our protocol. The database searches uncovered the following numbers of studies: Embase, 1421; IEEE, 288; Medline, 753; and PubMed, 1776. After discarding duplicates, 2880 studies remained. An additional four studies were uncovered by reading the references. One reviewer discarded 2770 papers after reviewing the titles and/or abstracts. One hundred and seven studies were excluded following assessment of the full text by two independent reviewers making consensus decisions. One paper was included after discussion with a third reviewer. Many studies retrieved from the medical databases searches were excluded due to a lack of computational analysis of bowel sounds. In parallel, many studies retrieved from the IEEE search were discarded due to a lack of clinical investigation, i.e. no GI diagnosis offered or no comparison with healthy controls. Several papers with the most sophisticated technology and pattern recognition analysis were excluded for simply outlining methods for identifying bowel sounds, or because the authors mimicked gastrointestinal conditions through the administration of drugs. Thus, ultimately only 14 studies were included in this review. The study selection process is detailed in Fig. 1.

A summary of the 14 included studies is given in Table 3. The studies cover a variety of target conditions, index tests, and study type.

The majority of studies were preliminary in nature, i.e. case control (multi-gate or two-gate studies) to test for an association between a bowel sound feature or features and a target condition. Thus, they typically examined the potential usefulness of bowel sound analysis in diagnosis rather than true measures of diagnostic accuracy. Four studies involved ROC analysis to determine test cut-off points and provided data on sensitivity and specificity. In three cases, this was derived from multi-gate data [13, 14, 23]. The fourth study had a prospective, blinded, cross-sectional (single-gate) design, undertaken at multiple centres, and we would expect that this gives a better indication of accuracy in the clinical setting in which the test will be applied [24].

In two studies, the diagnostic outcome was continuous rather than discrete [20, 21]. Here, the authors assessed the degree of correlation between a continuous variable (colon transit time) calculated by standard methods and values estimated from models derived from bowel sound features.

These two studies employed relatively sophisticated acoustic signal processing and modelling techniques. However, typically most studies involved only rudimentary bowel sound analysis techniques, and this is reflected in the simplicity of the analytical techniques employed. The duration of bowel sound recordings used was generally short, but ranged from 16 to 1 h. In addition, in most studies, the level of computational analysis was relatively low: simple signal processing rather than advanced pattern recognition. In only one case was there an attempt at localization of the bowel sounds.

Consensus decisions between two reviewers were reached for all sub-sections of the modified QUADAS tool for all articles included. The majority of studies were preliminary studies with case-control design, which typically are poorer in quality due to limited challenge bias and spectrum effects [27, 28]. These are grouped separately in Fig. 2. Generally, there was a high risk of bias in both the included DTA and preliminary studies.

Domain 1, patient selection, frequently gave rise to a high risk of bias because of a lack of information regarding study participant characteristics and inclusion/exclusion criteria. Three of the DTA studies had a case-control design. In addition, sample sizes were typically small which may have negatively impacted the reliability of results. For example, Kaneshiro et al.’s study [24] was well designed (blinded, prospective, longitudinal, single-gate (cohort) study across multiple centres, featuring a consecutive sample of patients), but it had only a small sample size of 28 participants, only nine of whom developed post-operative ileus.

There were no concerns about the applicability of the index test in any of the studies. However, we determined that it could have given rise to bias in over half of the preliminary studies and all the DTA studies. In all these, the bowel sound analysis was objective or, less frequently, was conducted blind [24] or prior [26]. However, in all cases, we determined that the index test could still have given rise to bias because the bowel sound feature or threshold was not pre-specified [13‐16, 18, 19, 21, 23, 24]. Tests of association or correlation between the target condition and a range of different features were made to find one of interest, or the cut-off threshold was determined as part of the study.

Domain 3, the reference standard, was considered to lead to a high risk of bias in three preliminary studies. In Craine et al.’s 1999 IBS focused paper [13], there was a lack of clarity on two counts, and ratings of unclears led to us reporting a high risk under our protocol. The reference standard was the Rome II criteria which lacks reliability in the absence of other investigations. It was also unclear whether the diagnosis was made without knowledge of the bowel sounds analysis results. Similarly, we rated Hadjileontiadis et al.’s [16] study as having a high risk of bias in this domain since it was unclear on reference standard reliability and timing relative to bowel sounds analysis. Ching and Tan’s [18] study on bowel obstruction was considered susceptible to a high level of bias because the reference standard was not objective and was made after the bowel sound analysis (diagnosis confirmed by clinical follow-up, by clinical evaluation, and by radiological and operative findings). There was an unclear risk of bias in the further three studies, where we could not determine if the diagnosis was made without knowledge of the bowel sound analysis or not [17, 19, 23]. This problem was rectified in the second study with the AGIS system [24] where there were different teams undertaking clinical assessment and bowel sound analysis.

In all but one case, there was no cause for concern about the applicability of the reference standard to the review question. There was concern that the target condition in the study by Liatsos et al. [26] small volume ascites as defined by the reference standard (without knowledge of patient comorbidities) is not solely a GI condition. Indeed, this was the one paper where the decision on inclusion in the review had to be made by the third reviewer (BJM).

Flow and timing were poorly described in two studies by Craine et al. of bowel sounds analysis in relation to IBS and other conditions [13, 15], leading to an unclear risk of bias. In a third study by the same group [14], not only was the interval between index test and reference standard test unclear, but it also appears that a mix of different methods was used to diagnose Crohn’s disease. Similarly, in four other studies [17‐19, 26], there was a variation in the reference standard used for subjects. In Tomomasa’s study [22], perhaps for ethical reasons, the healthy infant controls did not undergo a standard test to assess the rate of gastric emptying, and this introduced a risk of bias to the study.

Not all data was included in the analysis for two studies. In Ching and Tan’s study [18] of intestinal obstruction patients, six recordings were of poor quality and so were not included in the analysis, whilst the results from an IBS patient were not mentioned in the results section of the Hadjileontiadis et al. paper, which was also unclear on the relative timing of the reference standard and index tests.

The overall quality of statistics and reporting was disappointing for several studies. Problems related to statistics were found for two of Craine et al.’s studies [13, 15]. We determined that some of the statistics used in their 2002 study [15] were not appropriate. They looked for heterogeneity in sound to sound interval between groups, but since the non-ulcer dyspepsia groups were split based on this, the approach was circular. We were unable to determine if some of the statistics reported in an earlier Craine et al. paper [13] were appropriate since the test used was not detailed leading us to give them an unclear risk of bias in this domain.

The paper by Hadjileontiadis et al. [16] had quite sophisticated analysis but was still deemed to be at risk of bias due to statistics or reporting bias. No p values were provided, simply scatter plots and the IBS data was missing.

Four other studies had missing test outcomes [17, 21, 25]. Exact p values were missing in the two papers on delayed gastric emptying [20, 21], and one on acute abdomen [19].

The statistics on heterogeneity across groups for the first component of the first paper on the AGIS system [23] were considered to add bias, since the small sample sizes meant that non-parametric tests would have been more appropriate. This was rectified in their second paper, which detailed a similar method of bowel sound analysis for post-operative ileus diagnosis [24].

The DTA studies were all considered to have a high risk of statistical bias since the statistics provided on measures of accuracy were all derived from the same data from which the cut-off thresholds were determined, rather than from independent cases.

The pattern observed in risk of bias in relation to competing interests largely reflects the date of publication of the studies. Absence of information about competing interests and/or funding led to us record high risk in this domain for the earlier publications. However, we believe this was largely because many journals have only recently required statements in these areas.

The heterogeneity in index tests and target conditions precluded statistical meta-analysis. However, we are able to present a narrative describing the study results organised by target condition and bowel sound analysis approach.

Gastrointestinal disorders are a common cause of morbidity worldwide. There is a need for non-invasive, simple, diagnostic tests to reduce the demand for gastroenterology referrals and for investigations such as endoscopy and imaging with their associated risks. Our aim was to assess the potential for BSCA to meet this need.

Our review had several strengths. We conducted a highly comprehensive search of both the medical and engineering literature, without limiting study design type or GI condition to gain a full understanding of the applicability of BSCA. To allow quality assessment of the breadth of studies uncovered, we developed a novel quality assessment tool with parallel question sets for DTA and other study types.

All 14 included articles assessed associations or correlations between bowel sound features and GI conditions, many with highly significant results. Four of the studies also incorporated preliminary studies of diagnostic accuracy. These DTA studies provide only a moderate level of evidence to support the idea that bowel sound analysis is currently useful as a tool in GI diagnosis. However, together all the studies provide excellent evidence that many GI conditions are characterised by specific bowel sound features. Hence, there is strong evidence of the potential for future use of BSCA in diagnosis of GI disease and disorders.

The main diagnostic benefits or strongest associations between a bowel sound feature and condition were demonstrated in papers examining a single GI condition (e.g. IBS [13], ascites [26]) or where there was estimation of a single variable such as colon transit time [21]. However, it is important to note that the vast majority of these studies were case-control in design which can lead to inflated accuracy measures (see below).

Interestingly, the data also seems clearest where the target condition is associated with motility, e.g. hypomotility in post-operative ileus [24], disordered motility in IBS [13], delayed gastric emptying in adults [21] and infants [22], or prediction of colon transit time.

Whilst clinical applicability increases with two gate studies utilising control patients with conditions causing similar symptoms, incorporating multiple pathologies and variables appeared to make BSCA less reliable, or at least more difficult. In these cases, it necessitated study of multiple features, or recordings from multiple sites. For example, multiple features are needed to differentiate between causes of acute abdomen [19] and results were mixed in studies of bowel obstruction [17, 18]. The latter is perhaps surprising since traditionally, auscultation has been used in diagnosis of bowel obstruction with an expectation that this condition results in distinctive higher pitched sounds with a tinkling quality (Talley and O'Connor [29]).

The clearest and most reliable evidence for the utility of BSCA comes from tests of the AGIS system in diagnosis of post-operative ileus. Sensitivity and specificity were not overly impressive, but there is promise of a high negative predictive value, which is key in the clinical setting [23, 24]. The NPV varies with the prevalence of the target condition. However, given this study was undertaken in the setting in which the test will be used, we can be hopeful about the applicability of the results. The latest published study [24] is based on a small sample size; the full results from the clinical trial are needed to have full confidence in its utility. However, it appears that the AGIS system will prove useful in both prognosis and diagnosis in the post-operative setting. Certainly, it is likely that the system will provide useful additional information for the clinician to use in decision-making, even if it is not used as a stand-alone test. This work is particularly exciting because it also demonstrates that monitoring is possible with long-term recordings and real-time feedback.

Evidence from more rudimentary DTA-type studies suggested that BSCA may be useful in the diagnosis of IBS and in the differentiation between functional gastrointestinal disorders [13, 15]. In particular, it appears that study of the s-s interval could be used in IBS diagnosis. However, BSCA with technology at the level of included studies is unlikely to remove the need for endoscopy and imaging, especially given BSCA alone was not effective in differentiation between IBS and IBD.

Several studies were excluded from our review for reasons including a lack of GI disease, non-English full text, or no abstract [30‐53]. Although of less value than included papers, some excluded papers add information to help us assess the feasibility of the overall approach and are worthy of discussion.

Some provide information specifically about the clinical possibilities for BSCA. Three studies were excluded because only their abstracts were in English; however, they appear to add weight to the argument for utility of BSCA in diagnosis of post-operative ileus [31‐33].

Two short letters were excluded from the review, due to lack of abstract, but have some relevance to understanding of the utility of Craine et al.’s technology in the diagnosis of IBS and IBD [34, 35]. Yuki et al. [34] using the Enterotach technology found no significant effect on s-s interval in response to administration of pro-kinetic drugs or a stimulant laxative to mimic bowel disease. The authors suggested that the short recording time may not be long enough to detect alterations in the bowel sounds. In response, Craine and Silva [35] suggested that the effects of drugs on bowel sounds may not be easily predicted. They suggested that the decreased s-s interval reflects disordered motility rather than increased motility and sited unpublished data showing that both diarrhoea-dominant and constipation-dominant IBS patients have markedly increased fasting rates of sound production. Yuki et al. [34] were also unable to differentiate between inflammatory bowel disease patients and normal controls using the Enterotach analysis system.

However, overall, there were very many excluded papers that support the idea that BSCA could be useful. Some studies included patients with GI disease but were excluded because there was no independent reference standard for diagnosis. For example, Yamaguchi et al. [36] looked at the bowel sounds of diabetic patients with delayed gastric emptying. They found a lower sound index for gastroduodenal sounds (sum of the amplitude) in diabetic patients after food intake when compared to controls. Similarly, Goto et al. [37] studied bowel sounds in patients with sepsis and found a feature that negatively correlated with interleukin-6 blood concentration. Ozawa et al. [38] found reduced bowel sounds in Parkinson’s disease and multiple system atrophy patients. These studies hint at the breadth of conditions BSCA could be applicable to in the future. Logic would also suggest that pathology affecting the gut lumen (e.g. luminal masses, stricturing disease, or diverticular disease) could also be diagnosed via BSCA, which could provide an alternative screening test for colorectal malignancy.Other studies excluded were those utilising drug administration as a mimic of GI disease. Tomomasa et al. [39] and Emoto et al. [30] delivered drugs known to affect bowel motility, e.g. cisapride, scopolamine, and mosapride, allowing the authors to document changes in bowel sounds in the post-prandial state. These studies prompt us to think it is possible that BSCA may also be useful in development of an objective monitoring tool to assess the effects of GI treatments or GI symptoms as patients undergo management. In addition, BSCA could be used as a tool to test for side effects on the GI tract of a whole range of medications.

Some excluded studies simply demonstrate that bowel sounds can be extracted from the wealth of sounds emanating from the abdomen and surrounding environment, processed, and analysed [40, 41]. These are vital first steps in the use of BSCA for diagnosis. Many of these papers were more recent and indicate advancement of technology and more sophisticated analysis, which may provide powerful results if ever applied to a clinical setting [40, 42‐46]. They certainly throw into stark relief the very simple analysis on short recordings undertaken in the older studies, such those by Craine’s group [13].

We note that increased computer processing power has allowed use of longer recordings [40], which could be useful in the long-term monitoring of conditions, use of data from multiple sensors [46, 47], and real-time analysis [48]. The more recent studies also utilise signal processing techniques that are much more sophisticated [44, 45, 50‐53]. In addition, we see exciting pattern recognition and machine learning techniques being employed [30, 46]. This approach is likely vital when needing to utilise large amounts of data, e.g. from longer recordings, multiple sensors, or multiple features.

Perhaps the most significant limitation on this review is the heterogeneity between studies preventing formal statistical analysis. For example, no two DTA studies employed exactly the same BSCA methodology with the same cut-off point being tested for a given target condition. Similarly, no two heterogeneity studies employed the same statistics to allow comparison.

Diversity of bias assessments for different articles warranted use of a heavily modified QUADAS tool. Whilst this adapted tool is not validated, it did include questions derived from QUADAS-2 and other checklists that have been validated. It would not have been possible to adequately assess these studies using a single existing validated tool because of the vast heterogeneity of included studies. Note, we still presented the results of the preliminary and DTA studies separately to highlight the general difference in quality.

We searched both the medical and engineering literature in order to undertake a comprehensive search but decided to only include peer-reviewed studies to ensure the reliability of studies. Hence, we did not search clinical trials registers. We gained advice on our search strategy but did not use the Peer Review of Electronic Search Strategies checklist and, due to limited human resources, only included studies in English.

Our screening strategy was limited in that only a single reviewer undertook initial screening, although in the event of uncertainty, a second person was involved.

A major limitation of the included studies themselves was that the majority were dated and only employed older techniques of analysis and recording of sound data. Other limitations of the included studies were revealed by the quality assessment process: small sample size, poor specification of reference standard used, and weak statistical analysis including preliminary tests for association. We are aware that we have recorded a high risk of bias for the majority of studies, which diminishes our confidence in the potential for BSCA usefulness. In part, this reflects the study design. Many were tests of association or case-control in design allowing comparison between two groups: healthy controls and individuals with well-developed disease status. Whilst this type of preliminary study is extremely useful to screen whether a diagnostic test is worth developing, it can lead to an overestimate of the likely sensitivity and specificity of the test when used in the clinical setting. Use of healthy controls leads to inflated estimates of specificity, since few will have other diagnoses that could generate false positives. Similarly, inclusion of individuals with advanced disease will generate fewer false negatives and hence increase estimates of sensitivity. It was refreshing to review the paper from Kaneshiro et al. with a more sophisticated single-gate (cohort) design and consecutive sampling.

However, in at least one domain (domain2: index test), we may have been too severe in our decision-making. Studies examining multiple features, rather than one pre-specified criterion, were assessed as being at high risk of bias [16, 20, 21]. Nevertheless, we are aware that in some cases, it was necessary to use multiple features to enable differential diagnoses through BSCA, and hence, there this proved a very effective approach.

Generally, there is a significant need for improved study design: including more cross-sectional (single-gate) DTA studies in the appropriate clinical setting. With time, multiple well-designed studies are necessary to allow for meta-analysis and greater confidence in our conclusions.

Whilst hopeful about future prospects, we cannot yet recommend any BSCA-based diagnostic test to clinicians. The majority of included studies gave significant results, but only tested prototype technology and were simple preliminary case-control studies to test for association, or correlation studies. Where technology is closer to market and study design was more advanced [24], we still found problems in study quality, especially related to small sample sizes and patient selection. As mentioned above, replication of high-quality studies and larger sample sizes are necessary. Hopefully, this is likely to be addressed, for the AGIS system at least, in the next few years.

Generally, we found that researchers have much work to do to before BSCA can be applied in clinical practice. However, there is a great opportunity for gastroenterologists to work with engineers and software developers to provide access to patients in the correct setting and inform on good trial design. Together such collaborations should be able to provide clinically relevant high quality data, and possibly real progress in GI diagnostics.