Canine olfaction combined with Bayesian modeling for multi-cancer detection from breath samples: a Phase-2 study in India

Canine olfaction combined with Bayesian modeling for multi-cancer detection from breath samples: a Phase-2 study in India | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Canine olfaction combined with Bayesian modeling for multi-cancer detection from breath samples: a Phase-2 study in India Sanjeev Kulgod , Basavaraj R. Patil , Shashidhar K , Rakesh S Ramesh , Kiran Kulkarni , Somashekhar S P , Swaratika Majumdar , Akshita Singh , Claire Guest , Rob Harris , Sahana Shanbhag , Ido Aviram , Itamar Bitan , View ORCID Profile Akash Kulgod doi: https://doi.org/10.1101/2025.09.21.25336259 Sanjeev Kulgod 1 RadOn Cancer Centre , Hubli, India MCh Find this author on Google Scholar Find this author on PubMed Search for this author on this site Basavaraj R. Patil 2 Karnataka Cancer Therapy and Research Institute , Hubli, India MS, FRCS Find this author on Google Scholar Find this author on PubMed Search for this author on this site Shashidhar K 3 Karnataka Medical Centre and Research Institute , Hubli, India MCh Find this author on Google Scholar Find this author on PubMed Search for this author on this site Rakesh S Ramesh 4 St John’s Medical College , Bangalore, India DNB Find this author on Google Scholar Find this author on PubMed Search for this author on this site Kiran Kulkarni 4 St John’s Medical College , Bangalore, India MCh Find this author on Google Scholar Find this author on PubMed Search for this author on this site Somashekhar S P 5 Narayana Health City , Bangalore, India MCh Find this author on Google Scholar Find this author on PubMed Search for this author on this site Swaratika Majumdar 6 Aster CMI Hospital , Bangalore, India DM Find this author on Google Scholar Find this author on PubMed Search for this author on this site Akshita Singh 6 Aster CMI Hospital , Bangalore, India DNB Find this author on Google Scholar Find this author on PubMed Search for this author on this site Claire Guest 7 Medical Detection Dogs , United Kingdom MSc Find this author on Google Scholar Find this author on PubMed Search for this author on this site Rob Harris 7 Medical Detection Dogs , United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Sahana Shanbhag 8 Dognosis India Pvt Ltd , Bangalore, India Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ido Aviram 8 Dognosis India Pvt Ltd , Bangalore, India Find this author on Google Scholar Find this author on PubMed Search for this author on this site Itamar Bitan 8 Dognosis India Pvt Ltd , Bangalore, India Find this author on Google Scholar Find this author on PubMed Search for this author on this site Akash Kulgod 8 Dognosis India Pvt Ltd , Bangalore, India Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Akash Kulgod For correspondence: akash{at}dognosis.tech Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract PURPOSE Low-cost, acceptable, and high-sensitivity triage tests are needed to address the challenge of low cancer prevalence in population screening, particularly in low- and middle-income countries (LMICs). Breath-based canine olfaction has the potential to serve this role; however, evidence to date has been mostly limited to high-income countries and relatively small, single-cancer studies. We evaluated the analytical validity of a multi-cancer breath detection system using trained dogs and Bayesian fusion modeling. PATIENTS AND METHODS We conducted an assessor-masked, multi-center case–control study across six hospitals in Karnataka, India (March 2024–June 2025; CTRI/2024/10/075938). A total of 3,275 participants were enrolled: 1,773 for training and 1,502 for testing. The test cohort comprised 283 treatment-naive, biopsy-confirmed cancer patients (seven major cancer groups) and 1,219 controls (healthy, non-oncologic chronic disease, and benign biopsy). Breath was collected on cotton masks, stored under −20°C cold-chain conditions, and presented on a sniffing platform to trained detection dogs. Individual responses were integrated with Bayesian fusion incorporating historical dog performance and participant-level sample variables. RESULTS The fusion system achieved 91.5% sensitivity (95% CI, 88.0 – 94.8) and 90.8% specificity (95% CI, 89.1 – 92.5), with an area under the ROC curve (AUC) of 0.962 (95% CI, 0.951 - 0.971). Sensitivity was 89.6% in early-stages (Stage I–II), and was relatively consistent across major cancer types. CONCLUSION In a 1,502-participant test cohort, canine olfaction–Bayesian fusion achieved high accuracy for multi-cancer detection from breath, with stable performance across stages. These data establish the analytical validity and support the prospective evaluation of true screening populations. Key Objective Does canine olfaction combined with Bayesian modeling maintain analytical validity for multi-cancer breath screening in a large assessor-masked study in India? Knowledge Generated Detection dogs achieved sensitivity and specificity above 90% (AUC 0.962) in 1502 participants, with comparable performance across early- and late-stage cancers. Introduction Multi-cancer early detection (MCED) tests 1 , 2 aim to address the persistent problem of late-stage cancer diagnosis, which lowers treatment success and contributes substantially to global cancer mortality (9.7 million deaths in 2022). 3 Despite their promise, current blood-based MCED approaches have limited sensitivity in early-stage disease 4 and require costly sequencing infrastructure, trained personnel, and cold-chain logistics, with per-test costs of approximately US$500-1,000. 5 These characteristics constrain population-scale deployment, especially in low- and middle-income countries (LMICs). India, an LMIC with a population exceeding 1.4 billion, faces rising cancer incidence and, for some malignancies (e.g., breast cancer), an earlier age at onset than high-income settings. 6 Screening coverage remains low (approximately 1% ever screened), and a high proportion of cancers are detected at stages III–IV (approximately 80%), underscoring the need for accessible and acceptable screening alternatives. 7 At the same time, India’s sheer population size and relatively lower age-standardized incidence compared with many high-income countries 3 make the problem of low prevalence in asymptomatic cohorts particularly acute. Even tests with very high specificity will generate large absolute numbers of false positives relative to true cancers, thus straining the already limited diagnostic resources. A central challenge for early detection is therefore statistical as well as logistical: in lowprevalence populations, even highly specific tests yield modest positive predictive value (PPV). 8 The CanTest Framework was developed to guide evaluation of new tests across the diagnostic pathway, emphasizing the need for inexpensive, acceptable, high-sensitivity “triage” tests that enrich pretest probability so that confirmatory diagnostics can be deployed more efficiently. 9 This logic is especially pertinent in LMICs such as India, where confirmatory investigations are costly, invasive, and concentrated in urban centers, and patient acceptance of invasive procedures is low. Against this backdrop, breath-based canine detection could serve as a first-layer triage: non-invasive, lowcost, and operationally simple, it could filter large populations so that confirmatory diagnostics are focused on a smaller, higher-risk subset. 10 Over two decades of studies have indicated that trained canines can detect cancer-associated volatile organic compounds (VOCs) emitted from blood, urine, saliva, sweat, feces, and breath. 11 – 13 Breath is especially attractive, literature reviews have catalogued over 1,400 distinct VOCs in human breath, capturing tumor-related metabolic signatures. 14 Prior studies have evaluated canine detection for single cancers (e.g., breast 15 and lung 16 , 17 ) and, more recently, multi-cancer settings, 18 often reporting high sensitivity and specificity. Because breath collection is non-invasive and has modest storage/transport needs relative to phlebotomy, and canine teams can be trained and maintained at a comparatively low cost, breath sampling may support scalable population-level multicancer triage, enabling more effective use of organ-specific screening resources. This study covers a Phase-2 9 evaluation of trained dogs for multi-cancer detection from breath collected on face masks in India. We conducted an assessor-masked, multi-center case–control evaluation to estimate accuracy across cancer types and stages and assessed whether a Bayesian fusion integrating individual-dog historical performance, participant/sample covariates, and sensorderived behavioral signals improves accuracy, stability, and reproducibility relative to individual dog reads. Methods Participants We conducted an assessor-masked, multi-center case–control diagnostic accuracy study at six hospitals in Karnataka, India : three in Hubballi (RadOn Cancer Center, Karnataka Cancer Therapy and Research Institute, Karnataka Medical College & Research Institute) and three in Bengaluru (St John’s Medical College, Narayana Health City, Aster CMI). The two Hubballi public/tertiary centers draw predominantly rural catchment populations, whereas one Hubballi private center and the Bengaluru centers serve largely urban/semi-urban populations, yielding broad demographic and exposure diversity. The study was registered under the Clinical Trials Registry of India (CTRI/2024/10/075938) Recruitment occurred across oncology and non-oncology services (e.g,, internal medicine, surgery, gastroenterology and gynecology). Case enrollment was opportunistic with no restriction on the primary tumor site (multi-cancer intent); principal investigators at each site, supported by clinical research coordinators, approached eligible patients in outpatient clinics, inpatient wards, and procedure areas. Recruitment for control participants also occurred at community screening camps conducted by partner hospitals. The patients were adults ( > 18 years old) with biopsy/FNAC-confirmed malignancy and no prior cancer. Breath collection was performed before treatment to avoid confounding from acute procedures or therapy. The majority of cases were recruited after confirmation of the diagnosis, although a small minority were collected before biopsy. The controls were also adults ( > 18 years) and comprised (i) healthy volunteers, (ii) patients with non-oncologic chronic conditions, and (iii) patients with benign biopsy results. To reduce setting-related confounding, where feasible, controls from hospitals were recruited in the same rooms as cases within the same clinical departments. Key exclusions for all participants were pregnancy/postpartum < 90 days and any communicable respiratory illness, the latter to ensure the bio-safety of all team members. All participants provided written informed consent in their preferred language (Kannada/Hindi/English). Site-specific ethics approvals was obtained from the institutional committees; and the study procedures adhered to the Declaration of Helsinki and ICH-GCP. For transparency, the protocol was registered with the Clinical Trials Registry of India in October 2024 (CTRI/2024/03/061847). Demographics, tobacco use (smoking/chewing), and comorbidities were recorded for all participants. For cases, tumor site, histology, and stage were extracted from the medical records. The dataset was prospectively partitioned into disjoint training and testing sets at the participant level and all samples used for testing were novel samples from novel individuals at the time of presentation to the dogs. Index Test and Procedures Exhaled breath volatile organic compounds (VOCs) collected on cotton face masks were evaluated by trained detection dogs using a standardized indication protocol. The primary endpoint was cancer vs non-cancer classification for each sample. For all tests, samples were identified only by QR codes; clinical data and reference results were withheld from the handler logging the canine indications. The reference standard for case/control status was histopathology or cytology from routine care (biopsy/FNAC); imaging was used for the clinical context only, and index-test results did not influence clinical management. Sample Collection (Pre-analytical Conditions) The participants were instructed to perform 10 minutes of tidal breathing wearing a standardized cotton surgical mask. Trained research coordinators handled masks with gloves to minimize contamination, immediately sealed them in aluminum barrier bags, and logged unique QR codes under chain-of-custody. All participants in this study provided only one sample. Short-term storage was at 4-8 ° C (typically < 24 hours during on-site collection), followed by site-storage for 3 weeks at −20 ° C, and eventually transferred to the central testing laboratory at −20 ° C. Collection procedures were identical for cases and controls and, where feasible, performed in the same clinical rooms to limit location- or workflow-related confounding. The samples were stored for up to six months before testing. Canine Team and Experimental Platform Seven adult dogs (four Beagles, one Labrador, one Labrador-Indie mix, and one Dutch Shepherd-Belgian Malinois mix; 6 female, 1 male; ages 1-3 years) underwent positive-reinforcement training (10 weeks preceding each of the two trial sessions) at Dognosis facilities. Dogs were housed with daily exercise, veterinary oversight, and environmental enrichment, meeting the Five Freedoms and adhering to international welfare frameworks. 19 , 20 The protocol was approved by the Vigilomics Institutional Animal Ethics Committee (IAEC) under CCSEA registration and complied with the Prevention of Cruelty to Animals Act (1960) and CCSEA guidelines. Training used 1,773 samples from unique participants (373 cancer, 1,400 control); a sample was not used more than three times during training, and the same samples were not used in the training and testing sets. Both training and testing employed a custom eight-port sniffing workstation equipped with synchronized video and infrared sensors for automated timing and behavioral capture, as described in detail elsewhere. 21 QR scanners automated sample identification and logging, barrier gates controlled access timing, an automated feeder delivered rewards. The sensor streams were synchronized using software. 22 Testing Procedure The testing sessions were pre-declared and logged daily. Arrays of eight samples were randomly selected from the testing pool, yielding 40-70 samples per day. The dogs were evaluated individually in sequential runs. Two trainers and two laboratory technicians conducted the sessions; the handler calling indications remained masked to the sample identity throughout. Each sample was independently assessed by a minimum of three dogs using a pre-specified indication protocol (binary alert vs pass). Trials affected by technical interruptions (e.g., hardware interruption) or discontinued by the handler due to a lack of canine motivation to work were re-run when feasible or excluded from the primary accuracy analysis. Bayesian Fusion Algorithm and Statistical Analysis Individual dog indications and sample/demographic co-variates were combined into a Bayesian fusion framework. Dog-specific likelihoods were estimated from training folds (historical sensitivity/specificity per dog), and co-variate informed prior odds were obtained from a logistic model fit on the training set. For a given sample 𝔖, let C ∈ {0, 1} denote the cancer status; Prior P 0 = P ( C = 1 | x 𝔖 ), [See Priors]. Dogs i = 1, …, D produce indications I i ∈ {0, 1} when present; each dog has sensitivity s i and specificity c i estimated from the training confusion matrices. For the presented dogs i ∈ 𝔻 (𝔖), using per-dog sensitivity s i and specificity c i : Assuming conditional independence across dogs, the likelihoods are: Optionally tempering the likelihoods with τ ≥ 0: The Bayesian update with the prior P 0 is: Posterior probabilities were compared with varying decision thresholds; threshold–performance trade-offs are shown as ROC/operating-point analyses. All fusion computations were performed on a test set without re-estimation of the dog performance parameters. The primary accuracy measures were sensitivity, specificity, and area under the ROC curve (AUC) for the testing set with two-sided 95% CIs. Pre-specified stratification included stage (I–II vs III–IV), cancer type, sex, age, tobacco use, tumor site, and dog identity. Prior Prediction Model Feature Engineering and Preprocessing Engineered variables included age, sex, tobacco exposure, family history of cancer, self-reported persistent symptoms, and other variables. Numeric variables were imputed and standardized, while categorical variables are imputed and one-hot encoded. For sample x 𝔖 , let the features be ϕ ( x 𝔖 ) Classifier A supervised classifier (logistic regression) was trained on training historical data to produce P 0 ( x ) = P ( C =1 | x 𝔖 ). For the logistic regression, where ϕ ( x 𝔖 ) is the feature vector of engineered covariates for sample x 𝔖 , β is the vector of regression coefficients, and β 0 is the intercept term. Results Participants A total of 3,625 individuals were assessed across the six hospitals for eligibility. After 350 exclusions (age, communicable respiratory disease, and pregnancy); 3,275 participants were enrolled. Of these, 1,773 contributed to the training cohort and 1,502 to the testing cohort. Within the control group in the test set, participants were comprised of healthy volunteers (n = 789), patients with chronic non-malignant conditions (n = 426), and individuals with benign biopsy findings (n = 4). The final test cohort comprised 283 biopsy-confirmed cancers and 1,219 controls, with a STARD flow diagram summarizing the recruitment ( Figure 1 ). Download figure Open in new tab Figure 1. STARD flow diagram of participant disposition showing exclusions, training allocation, and test cohort composition. Cases tended to be older than controls (mean 55.3 vs 48.5 years) and had higher prevalence tobacco use, particularly chewing (38.2% vs 14.9%). Smoking history was also more common among the cases (14.1% vs 7.6%). However, given the larger number of controls overall (n = 1,219), the absolute number of tobacco-exposed individuals was comparable between the groups (e.g., 182 controls with chewing history vs 108 cases). This balance reduces the likelihood of canine discrimination reflecting tobacco exposure. Hypertension and diabetes were similarly distributed among the groups ( Table 1 ). No adverse events occurred during breath collection or testing. View this table: View inline View popup Download powerpoint Table 1: Baseline demographic and clinical characteristics of the test cohort (n = 1,502) Cancer distribution The distribution of the cancer groups in the test cohort (n = 284 cancer diagnoses from 283 participants) is summarized in Table 2 . One participant had two distinct primary cancers, both of which were included in the overall count but only one of them (Breast cancer) was used to determine performance. Grouping rules (site and histology mapping) are detailed in Supplementary Methods S2, with the full ICD-10/ICD-O mapping in Table S1 . Head and neck cancers accounted for nearly one-third of the cases, consistent with the high prevalence of tobaccorelated oral malignancies in India. Upper gastrointestinal and breast cancers together contributed to another 40% of cases, reflecting the dual burden of lifestyle-associated and hormonally driven malignancies in the region. Gynecological cancers were also well represented, while lung and male genitourinary cancers were fewer in number, paralleling the local incidence trends. View this table: View inline View popup Download powerpoint Table 2: Cancer-group distribution in the test cohort (n=284) Primary accuracy Bayesian fusion achieved an overall 91.5% sensitivity (95% CI, 88.0 – 94.8) and 90.8% specificity (95% CI, 89.1 – 92.5), with an area under the ROC curve (AUC) of 0.962 (95% CI, 0.951 - 0.971) ( Figure 2 ). At a probability threshold of 0.5, which was used as the default operating point for binary classification, the model correctly identified 259 of 283 cancers and 1107 of 1,219 controls, with 24 false negatives and 112 false positives ( Table 3 ). As expected, alternative thresholds yielded predictable trade-offs between sensitivity and specificity, captured in the ROC profile; full sensitivity/specificity curves across thresholds are shown in Figure 3 ). Positive and negative predictive values are not reported given the case–control design and artificially inflated prevalence, but will require evaluation in prospective screening populations. View this table: View inline View popup Download powerpoint Table 3: Confusion matrix, test set Download figure Open in new tab Figure 2. ROC curve for the Bayesian fusion (AUC = 0.962) Download figure Open in new tab Figure 3. Sensitivity and Specificity across decision thresholds Performance by cancer type The point-estimate sensitivity was high across all major cancer groups, ranging from 92.0% for head and neck cancers to over 94% for upper and lower gastroin-testinal cancers ( Table 4 ). Sensitivity remained > 87% in the four largest groups (head and neck, breast, upper GI, gynecologic), supporting consistent detection across diverse tumor types. The apparent 100% sensitivities for lung/thoracic, and male genitourinary cancers reflect small sample sizes (n < 10) and are accompanied by wide confidence intervals, indicating that these estimates should be interpreted cautiously. The “Others” category, which pooled rare cancers, showed wider variability with a sensitivity of 76.9% (95% CI: 53.8–100.0). The test set sensitivity reached 92.2% for all cancer groups excluding the “Others” category, compared to an overall sensitivity of 91.5% when all groups were included; the controls were not subdivided by cancer group. Taken together, these findings suggest that canine detection performance was robust across the cancer types most prevalent in India (head and neck, gastrointestinal, and breast cancer), without evidence of collapse in rarer categories, although precision in smaller groups will require larger cohorts. View this table: View inline View popup Download powerpoint Table 4: Sensitivity by cancer group in the test cohort Performance by stage The stage-specific sensitivity remained consistently high across all stages, 87.1% in stage I, 90.8% in stage II, 93.7% in stage III, and 91.5% in stage IV ( Table 5 ). Aggregated across stages, early-stage disease (I–II) was detected with a sensitivity of 89.6%, comparable to 92.5% for late-stage disease (III–IV). Confidence intervals overlapped across stages, and estimates for the “unknown” stage group (n = 14) were imprecise due to the small sample size. View this table: View inline View popup Download powerpoint Table 5: Sensitivity by cancer stage Subgroup-analysis Performance was broadly consistent across the demographic and exposure strata ( Table 6 ). The sensitivity was somewhat higher in participants aged > 50 years (94.8%) than in those < 50 years (85.9%), whereas the specificity was higher in younger individuals (95.1% vs 85.0%). Male and Female participants showed comparable sensitivity (90.6% vs 92.2%) and specificity (89.2% vs 91.7%), robust against differences in cancer type distribution (e.g., breast and gynecologic vs head and neck). Importantly, detection performance was resilient in both nevertobacco users (sensitivity 90.6%) and current/former users (92.7%), with overlapping confidence intervals, suggesting that discrimination was not solely attributable to tobacco-related exposures. View this table: View inline View popup Download powerpoint Table 6: Subgroup performance Robustness to Individual Dog Variation Although all dogs were trained under an identical protocol, the individual-level sensitivity and specificity varied ( Figure S1 and Figure S2 ). Because such variability can raise concerns regarding replicability and standardization in biomedical detection, we evaluated the robustness of the ensemble using two stress tests. Leave-one-dog-out We conducted an ablation analysis in which, for each of the seven dogs, the dog was removed from the ensemble and the Bayesian fusion model was re-evaluated without its indications. This procedure yields seven “reduced” ensembles, each reflecting performance if a given dog were unavailable. The results are summarized in Table 7 . View this table: View inline View popup Download powerpoint Table 7: Ablation study: Leave-One-Dog-Out “Bad-day” simulation To model the transient degradation in one dog’s signal at the session level, we performed a permutation stress test. For each testing session, we randomly selected one active dog and permuted the dog’s indications across the session’s samples, preserving all other dogs’ indications and the ground-truth labels. This destroys the association between the dog’s responses and the case/control status while preserving its marginal response rate. We repeated the procedure five times with distinct random seeds and reported the mean area under the ROC curve (AUC) across the replicates ( Fig. 4 ). Download figure Open in new tab Figure 4. Robustness of our model against individual dog variation Interpretation Across both analyses, ensemble performance remained stable relative to the full model, indicating that the fusion approach is resilient to (i) the removal of any single dog and (ii) temporary within-session degradation in one dog’s indications. This supports the robustness of multi-dog Bayesian fusion to idiosyncratic behavior at the individual-dog level. EHR-Derived Prior Contribution We quantified the relative contribution of electronic health record (EHR)-derived covariates to the prior P 0 using standardized contribution scores (percent importance, averaged across crossvalidated folds). The largest effects were observed for tobacco exposure (chewing+smoking) (29.50%), persistent symptoms (14.45%) and age group (8.02%), followed by sex (6.33%), firstdegree family history of cancer (6.28%) and comorbidities (5.32%)( Table 8 ). The directions and relative magnitudes align with established epidemiology — such as risk increasing with age and cumulative tobacco exposure and first-degree family history elevatating baseline risk. Incorporating these predictors (including the composite tobacco measure and an age × medication interaction) improved discrimination of the prior model by ΔAUC = +0.11 versus reduced specifications. Contribution ranks were stable under resampling, supporting the face validity and robustness of the priors and their added value when fused with canine indications. View this table: View inline View popup Download powerpoint Table 8: Prior contribution Ablation: Contribution of model components Figure 5 compares three operating variants on the test set: (i) Only Prior , which uses the EHR–derived covariate model without dog indications; (ii) Pack Majority , an unweighted majority vote across dogs that ignores covariates; and (iii) full Bayesian Fusion , which combines per-dog likelihoods with the prior. Discrimination improved from an AUC of 0.832 (Only Prior) to 0.932 with Pack Majority (Δ= + 0.100), and further to 0.962 with Bayesian Fusion (Δ= + 0.030 vs. majority vote; Δ= + 0.130 vs. prior). These results indicate that canine indications provide the dominant signal, whereas covariate-informed prior adds incremental, complementary information when fused probabilistically, thereby yielding a modest but meaningful gain over naive voting and producing well-defined posterior probabilities for thresholding. Download figure Open in new tab Figure 5. Ablation : Performance of Model Components Discussion India’s National Program for Prevention and Control of Non-Communicable Diseases (NP-NCD) operationalizes population-based screening for three cancers, oral, cervical, and breast, delivered through Health and Wellness Centers by healthcare workers using oral visual examination, VIA, and clinical breast examination, respectively. 23 Despite national scale-up, screening coverage remains modest, 24 and these three cancers together account for only about one-third of the incident cancers in India. 3 In low-prevalence populations, even highly specific tests yield a modest positive predictive value (PPV), amplifying the downstream workload. 8 A non-invasive, acceptable, lowcost triage test at the front end of NP-NCD pathways could enrich pretest probability and focus confirmatory diagnostics on a smaller, higher-risk subset—an approach aligned with the CanTest framework for evaluating tests across real-world diagnostic pathways. 9 Against this backdrop, we evaluated the analytical validity of multi-cancer breath-based canine detection. In a multi-center, assessor-masked case–control study across six Indian hospitals, a Canine–Bayesian fusion model demonstrated strong analytical validity (ROC AUC = 0.962) with balanced sensitivity and specificity above 90%. Sensitivity remained above 87% across stages I to IV, and performance was broadly consistent across major cancer groups, with the Bayesian fusion model mitigating between-dog variability relative to individual reads. These data support the potential role of breath-based triage in expanding participation and improving the efficiency of downstream diagnostics within workflows such as NP-NCD guidelines. However, prospective, population-based studies are needed to calibrate operating thresholds and quantify clinical utility under true screening prevalence. Breath-based detection interrogates volatile metabolic signatures and is therefore mechanistically complementary to blood-based multi-cancer early detection (MCED) assays that rely on circulating genomic/epigenomic markers. Our results suggest integrated pathways that can be evaluated across settings: (i) breath-first triage to enrich pretest probability before organ-specific imaging or blood-based MCED; (ii) reflex testing (breath → MCED or MCED → breath) to improve PPV while controlling downstream workload; and (iii) probabilistic fusion of breath and blood signals for a single operating point tailored to local resources. Such designs could increase reach (primary care, workplace, or home-based breath collection), improve acceptability, and reduce false-positive results. Head-to-head and combined-pathway studies—incorporating cost, throughput, and qualitycontrol metrics will be critical for determining where breath-based triage adds the greatest value in both high- and low-resource health systems. However, important limitations of this study remain. First, the case–control design is susceptible to spectrum and selection biases and typically inflates apparent performance relative to true screening settings. Second, the cancer prevalence in the test cohort ( ∼ 1:4) exceeds that of the general screening population, which would depress the PPV at low prevalence despite high sensitivity. Third, generalizability beyond participating centers cannot be stated with certainty, all sites were within a single Indian state (Karnataka), and unmeasured geographic, environmental, dietary, or microbiome factors may shift breath VOC backgrounds. Finally, batch effects and residual confounding were possible despite standardized procedures. Planned methodological enhancements To further reduce between-dog variability and strengthen the data-driven framework, we plan to augment the fusion with (i) computer-vision features from multi-view videos to estimate features such as dwell time, head pose, re-approach counts, and micro-pauses extracted via pose-estimation and event-detection models, and (ii) non-invasive physiological signals, including wearable EEG to capture brain activity before and during sniffing. In a hierarchical Bayesian formulation, dog-specific likelihood terms can be dynamically weighted by these signals and by contemporaneous performance calibration, yielding time-varying reliability weights for each dog–trial. This refinement is expected to narrow the uncertainty around the persample posterior probabilities, attenuate variable behavior, improve repeatability across sessions and sites, and improve the overall performance of the system. Prospective evaluation These results establish analytical validity but require confirmation in prospective, population-based studies to determine the clinical utility of true screening prevalence. Such studies should pre-specify thresholds and referral pathways and evaluate key endpoints (positive predictive value, detection rate per 1,000 screened, stage distribution at diagnosis, and time to diagnostic resolution) while monitoring downstream diagnostic yield, false-positive cascades, and participant acceptability. Integration within standard-of-care pathways and assessments of throughput, workflow, and costs will be essential but are beyond the scope of the present case–control study. In summary, breath-based multi-cancer detection using trained dogs augmented by Bayesian fusion demonstrated a strong analytical performance across cancer types and stages. The next phase of methodologically enriched modeling with EEG and machine vision, along with prospective evaluation and health-system integration, will be critical for determining clinical utility and implementation feasibility. Data Availability All data produced in the present study are available upon reasonable request to the authors Author Contributions Conception and design: Akash Kulgod, Itamar Bitan, Sanjeev Kulgod, Sahana Shanbhag, Claire Guest, Rob Harris Provision of study materials or patients: Sanjeev Kulgod, Basavaraj R. Patil, Shashidhar K., Rakesh S. Ramesh, Kiran Kulkarni, Somashekhar S. P., Swaratika Majumdar, Akshita Singh Collection and assembly of data: Sahana Shanbhag, Ido Aviram, Itamar Bitan, Akash Kulgod Data analysis and interpretation: Akash Kulgod, Itamar Bitan, Sahana Shanbhag, Ido Aviram Manuscript writing: Akash Kulgod (drafting); all authors (review and editing) Final approval of manuscript: All authors All authors are accountable for all aspects of the work Authors’ Disclosures of Potential Conflicts of Interest The following represents the disclosure of information provided by the authors of this manuscript. All relationships were considered compensated unless otherwise noted. Relationships are self-held unless otherwise noted. I = Immediate Family Member, Inst = My Institution. Relationships may not be related to the subject matter of this study. Sanjeev Kulgod, MCh Employment: RadOn Cancer Centre, Hubli. Role: Lead Principal Investigator. No other conflicts to disclose Basavaraj R. Patil, MS, FRCS Employment: Karnataka Cancer Therapy and Research Institute, Hubli. No other conflicts to disclose . Shashidhar K ., MCh Employment: Karnataka Medical Centre and Research Institute, Hubli. No other conflicts to disclose. Rakesh S. Ramesh, DNB Employment: St John’s Medical College, Bangalore. Research Funding: Dognosis India Pvt. Ltd (Inst). Kiran Kulkarni, MCh Employment: St John’s Medical College, Bangalore. Research Funding: Dognosis India Pvt. Ltd (Inst). Somashekhar S. P ., MCh Employment: Aster CMI Hospital, Bangalore. Research Funding: Dognosis India Pvt. Ltd (Inst). Swaratika Majumdar, DM Employment: Narayana Health City, Bangalore. No other conflicts to disclose . Akshita Singh, DNB Employment: Narayana Health City, Bangalore. No other conflicts to disclose . Claire Guest Employment: Medical Detection Dogs. No other conflicts to disclose . Rob Harris Employment: Medical Detection Dogs. No other conflicts to disclose . Sahana Shanbhag Employment: Dognosis India Pvt. Ltd. Stock and Other Ownership Interests: Dognosis Inc (ESOPs). Ido Aviram Employment: Dognosis India Pvt. Ltd. Stock and Other Ownership Interests: Dognosis Inc (ESOPs). Itamar Bitan Employment: Dognosis India Pvt. Ltd. Stock and Other Ownership Interests: Dognosis Inc. Akash Kulgod Employment: Dognosis India Pvt. Ltd. Stock and Other Ownership Interests: Dognosis Inc. Supplementary Material View this table: View inline View popup Table S1: Categorization of cancer types View this table: View inline View popup Download powerpoint Table S2: Baseline characteristics, training cohort (n=1,773) View this table: View inline View popup Download powerpoint Table S3: Cancer-group distribution in the training cohort (n=373) View this table: View inline View popup Download powerpoint Table S4: Distribution of samples across locations in training and test cohorts View this table: View inline View popup Download powerpoint Table S5: Distribution of comorbid conditions in training and test cohorts View this table: View inline View popup Download powerpoint Table S6: Specificity & Sensitivity by location of the sample Download figure Open in new tab Figure S1: Sensitivity and specificity for each dog. Download figure Open in new tab Figure S2: Confusion matrices for each dog. Acknowledgments We are deeply grateful to all patients who volunteered and provided breath samples for this study. We gratefully acknowledge the contributions of clinical partners across all sites involved in the BreathEasy study. At Radon Cancer Center, Hubli, we thank Dr. Sheetal Kulgod for leading the screening camps and Dr. Vishal Malvade for patient recruitment support. At KMCRI, Hubli, we thank Dr. Laxmikant, Dr. Manisha,, Dr Vishal Kulkarni, Dr Sunita Vernekar, Dr Rajesh H, Dr Parvati Jugular, Dr Kanchana, Nurse Ashwini, Manjunath and Chandrashekar for their support in recruitment. At KCTRI, Hubli, we acknowledge Dr. Umesh Hallikeri and Dr. Vilas Gunda for facilitating site access and screening camps, along with Dr. Saikumari and Dr. Poornachandra for their contributions to participant recruitment. We also extend our thanks to Dr. Nikhil, Bhagya and Janaki Sisters at St. John’s Medical College, Bangalore; Dr. Uday Shivaji, Dr. Karthik, Dr Anjum and Veena at Narayana Health City, Bangalore; and Mr. Surinder Kher and Ms. Rajeshwari at Aster CMI Hospital, Bangalore for site-level clinical operations support. We also thank Dr Shashikant Kulgod and Dr Akash Gadgade for their support throughout the study. We sincerely thank our Clinical Research Coordinators for their on-ground efforts and dedication to sample and data collection : Kiran B Nagashettikoppa, Rudresh Hiremath, Dr. Arnab Sarkar, Dr. Basavaraj Varad, Akash Sudi, Dr. Tahameen Kurubanavar, Dr. Anjum Jangliwale, Dasmen Pratiban, Dr. Kishore and Dr. Shabin. We also thank Dr Minal Dakhave for her support throughout her tenure. We thank the Dognosis lab team — Richard King, Manish, Gokul, Nirmal, and Venkat Rao — for their dedicated work managing canine operations. We recognize the contributions of the Dognosis R&D team, including Aswin R, Thirumalai Srinivasan, Abhijit Kadalli, Vishnu Kumar, Sree Subha Ramaswamy, Dilawar Singh, Maadesh Kumar, Subham Hota, Achin Parashar, Anushree Narjala, Shashikant Gadiwaddar, and Promit Moitra for their efforts in hardware development, data management, analysis, and model building. This study was supported by Dognosis India Pvt. Ltd. Footnotes Research support / funding: Dognosis India Pvt. Ltd. Prior presentation: Portions of this study (interim analysis; test n=200 samples) were presented at the ASCO 2025 (June 2025). Disclaimers: None. References [1]. ↵ Ruth Etzioni , Roman Gulati , and Noel S Weiss . Multicancer Early Detection: Learning From the Past to Meet the Future . JNCI: Journal of the National Cancer Institute , 114 ( 3 ): 349 – 352 , March 2022 . OpenUrl CrossRef PubMed [2]. ↵ Carmen E. Guerra , Prateek V. Sharma , and Brenda S. Castillo . Multi-Cancer Early Detection: The New Frontier in Cancer Early Detection . Annual Review of Medicine , 75 ( 1 ): 67 – 81 , January 2024 . OpenUrl PubMed [3]. ↵ Freddie Bray , Mathieu Laversanne , Hyuna Sung , Jacques Ferlay , Rebecca L. Siegel , Isabelle Soerjomataram , and Ahmedin Jemal . Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries . CA: a cancer journal for clinicians , 74 ( 3 ): 229 – 263 , 2024 . OpenUrl CrossRef PubMed [4]. ↵ E.A. Klein , D. Richards , A. Cohn , M. Tummala , R. Lapham , D. Cosgrove , G. Chung , J. Clement , J. Gao , N. Hunkapiller , A. Jamshidi , K.N. Kurtzman , M.V. Seiden , C. Swanton , and M.C. Liu . Clinical validation of a targeted methylation-based multi-cancer early detection test using an independent validation set . Annals of Oncology , 32 ( 9 ): 1167 – 1177 , September 2021 . OpenUrl CrossRef PubMed [5]. ↵ Clare Turnbull , Nicholas Wald , Richard Sullivan , Paul Pharoah , Richard S. Houlston , Ajay Aggarwal , Stuart Hogarth , and Margaret McCartney . GRAIL-Galleri: Why the special treatment? The Lancet , 403 ( 10425 ): 431 – 432 , February 2024 . OpenUrl [6]. ↵ Ravi Mehrotra and Kavita Yadav . Breast cancer in India: Present scenario and the challenges ahead . World Journal of Clinical Oncology , 13 ( 3 ): 209 – 218 , March 2022 . OpenUrl PubMed [7]. ↵ Kunal Oswal , Kumar Gaurav , Akash Pradhan , Venkataramanan Ramachandran , Paul Sebastian , and Partha Basu . Current Status of Implementation of Cancer Screening Programme in India: A Review of Policies and Practice . Asian Pacific Journal of Cancer Prevention : APJCP , 26 ( 4 ): 1117 – 1128 , 2025 . OpenUrl PubMed [8]. ↵ Mariska M. G. Leeflang , Patrick M. M. Bossuyt , and Les Irwig . Diagnostic test accuracy may vary with prevalence: Implications for evidence-based diagnosis . Journal of Clinical Epidemiology , 62 ( 1 ): 5 – 12 , January 2009 . OpenUrl CrossRef PubMed Web of Science [9]. ↵ Fiona M. Walter , Matthew J. Thompson , Ian Wellwood , Gary A. Abel , William Hamilton , Margaret Johnson , Georgios Lyratzopoulos , Michael P. Messenger , Richard D. Neal , Greg Rubin , Hardeep Singh , Anne Spencer , Stephen Sutton , Peter Vedsted , and Jon D. Emery . Evaluating diagnostic strategies for early detection of cancer: The CanTest framework . BMC Cancer , 19 ( 1 ): 586 , December 2019 . OpenUrl CrossRef PubMed [10]. ↵ Deb Schrag , Tomasz M Beer , Charles H McDonnell , Lincoln Nadauld , Christina A Dilaveri , Robert Reid , Catherine R Marinac , Karen C Chung , Margarita Lopatin , Eric T Fung , and Eric A Klein . Blood-based tests for multicancer early detection (PATHFINDER): A prospective cohort study . The Lancet , 402 ( 10409 ): 1251 – 1260 , October 2023 . OpenUrl [11]. ↵ Carolyn M Willis , Susannah M Church , Claire M Guest , W Andrew Cook , Noel McCarthy , Anthea J Bransbury , Martin R T Church , and John C T Church . Olfactory detection of human bladder cancer by dogs: proof of principle study . BMJ , 329 ( 7468 ): 712 , September 2004 . OpenUrl Abstract / FREE Full Text [12]. Claire Guest , Rob Harris , Karen S. Sfanos , Eva Shrestha , Alan W. Partin , Bruce Trock , Leslie Mangold , Rebecca Bader , Adam Kozak , Scott Mclean , Jonathan Simons , Howard Soule , Thomas Johnson , Wen-Yee Lee , Qin Gao , Sophie Aziz , Patritsia Maria Stathatou , Stephen Thaler , Simmie Foster , and Andreas Mershin . Feasibility of integrating canine olfaction with chemical and microbial profiling of urine to detect lethal prostate cancer . PLOS ONE , 16 ( 2 ): e0245530 , February 2021 . OpenUrl CrossRef PubMed [13]. ↵ Paula Jendrny , Friederike Twele , Sebastian Meller , Albertus Dominicus Marcellinus Erasmus Osterhaus , Esther Schalke , and Holger Andreas Volk . Canine olfactory detection and its relevance to medical detection . BMC Infectious Diseases , 21 : 838 , August 2021 . [14]. ↵ Natalia Drabińska , Cheryl Flynn , Norman Ratcliffe , Ilaria Belluomo , Antonis Myridakis , Oliver Gould , Matteo Fois , Amy Smart , Terry Devine , and Ben De Lacy Costello . A literature survey of all volatiles from healthy human breath and bodily fluids: The human volatilome . Journal of Breath Research , 15 ( 3 ): 034001 , April 2021 . OpenUrl [15]. ↵ Michael McCulloch , Tadeusz Jezierski , Michael Broffman , Alan Hubbard , Kirk Turner , and Teresa Janecki . Diagnostic accuracy of canine scent detection in early- and late-stage lung and breast cancers . Integrative Cancer Therapies , 5 ( 1 ): 30 – 39 , March 2006 . OpenUrl CrossRef PubMed Web of Science [16]. ↵ Klaus Hackner , Peter Errhalt , Michael Rolf Mueller , Manulea Speiser , Beatrice A Marzluf , Andrea Schulheim , Peter Schenk , Johannes Bilek , and Theodor Doll . Canine scent detection for the diagnosis of lung cancer in a screening-like situation . Journal of Breath Research , 10 ( 4 ): 046003 , September 2016 . OpenUrl PubMed [17]. ↵ Margaret A Crawford , Catherina L Chang , Sandra Hopping , Clare M Browne , and Timothy L Edwards . Influences of breath sample re-use on the accuracy of lung cancer detection dogs . Journal of Breath Research , 17 ( 1 ): 016001 , 2023 . OpenUrl [18]. ↵ Elizabeth Half , Adelina Ovcharenko , Ronit Shmuel , Sharon Furman-Assaf , Milana Avdalimov , Assaf Rabinowicz , and Nadir Arber . Non-invasive multiple cancer screening using trained detection canines and artificial intelligence: A prospective double-blind study . Scientific Reports , 14 ( 1 ): 28204 , November 2024 . OpenUrl PubMed [19]. ↵ UK Parliament . Animal welfare act 2006: Section 9 — duty of person responsible for animal to ensure welfare . https://www.legislation.gov.uk/ukpga/2006/45/section/9 , 2006 . xUK Public General Acts. [20]. ↵ European Union . Directive 2010/63/eu of the european parliament and of the council of 22 september 2010 on the protection of animals used for scientific purposes . https://eur-lex.europa.eu/eli/dir/2010/63/oj/eng , 2010. Official Journal of the European Union L 276, 20.10.2010, pp. 33 – 79 . [21]. ↵ Aswin R , Sai Krishna Teja Sadhu , Itamar Bitan , and Akash Kulgod . Towards Enactivist ACI - Sensor-Rich Olfactory Workstation and Suit for Detection Dogs . In Proceedings of the International Conference on Animal-Computer Interaction, ACI ‘24 , pages 1 – 9 , New York, NY, USA , December 2024 . Association for Computing Machinery. [22]. ↵ Christian Kothe , Seyed Yahya Shirazi , Tristan Stenner , David Medine , Chadwick Boulay , Matthew I. Grivich , Fiorenzo Artoni , Tim Mullen , Arnaud Delorme , and Scott Makeig . The lab streaming layer for synchronized multimodal recording . Imaging Neuroscience , 3 :IMAG.a.136, September 2025 . [23]. ↵ Ministry of Health and Family Welfare, Government of India . National programme for prevention and control of non-communicable diseases (np–ncd): Operational guidelines (2023– 2030) . Ministry guidelines , 2024 . Accessed 9 Sep 2025 . [24]. ↵ Sivakami Sundari Subramanian , Binukumar Bhaskarapillai , and Radhakrishnan Jayakrishnan . Female cancer screening in India: Results from the National Family Health Survey (NFHS-5) and the way forward . Journal of Family Medicine and Primary Care , 13 ( 10 ): 4552 – 4559 , October 2024 . OpenUrl View the discussion thread. Back to top Previous Next Posted September 22, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Canine olfaction combined with Bayesian modeling for multi-cancer detection from breath samples: a Phase-2 study in India Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Canine olfaction combined with Bayesian modeling for multi-cancer detection from breath samples: a Phase-2 study in India Sanjeev Kulgod , Basavaraj R. Patil , Shashidhar K , Rakesh S Ramesh , Kiran Kulkarni , Somashekhar S P , Swaratika Majumdar , Akshita Singh , Claire Guest , Rob Harris , Sahana Shanbhag , Ido Aviram , Itamar Bitan , Akash Kulgod medRxiv 2025.09.21.25336259; doi: https://doi.org/10.1101/2025.09.21.25336259 Share This Article: Copy Citation Tools Canine olfaction combined with Bayesian modeling for multi-cancer detection from breath samples: a Phase-2 study in India Sanjeev Kulgod , Basavaraj R. Patil , Shashidhar K , Rakesh S Ramesh , Kiran Kulkarni , Somashekhar S P , Swaratika Majumdar , Akshita Singh , Claire Guest , Rob Harris , Sahana Shanbhag , Ido Aviram , Itamar Bitan , Akash Kulgod medRxiv 2025.09.21.25336259; doi: https://doi.org/10.1101/2025.09.21.25336259 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Oncology Subject Areas All Articles Addiction Medicine (567) Allergy and Immunology (863) Anesthesia (297) Cardiovascular Medicine (4411) Dentistry and Oral Medicine (443) Dermatology (380) Emergency Medicine (606) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1505) Epidemiology (15205) Forensic Medicine (30) Gastroenterology (1119) Genetic and Genomic Medicine (6574) Geriatric Medicine (666) Health Economics (994) Health Informatics (4511) Health Policy (1365) Health Systems and Quality Improvement (1608) Hematology (537) HIV/AIDS (1263) Infectious Diseases (except HIV/AIDS) (15903) Intensive Care and Critical Care Medicine (1103) Medical Education (620) Medical Ethics (144) Nephrology (666) Neurology (6573) Nursing (345) Nutrition (998) Obstetrics and Gynecology (1139) Occupational and Environmental Health (954) Oncology (3319) Ophthalmology (968) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (662) Pediatrics (1689) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5422) Public and Global Health (9205) Radiology and Imaging (2191) Rehabilitation Medicine and Physical Therapy (1367) Respiratory Medicine (1191) Rheumatology (593) Sexual and Reproductive Health (709) Sports Medicine (529) Surgery (709) Toxicology (99) Transplantation (288) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9feaec5e7fae41e2',t:'MTc3OTI3NTcyNA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();