An Indian study has shown combining dogs’ noses with Bayesian fusion modelling is an accurate way to detect cancer on a patient’s breath, but experts feel there are more important lessons to learn from the findings.
Sniffer dogs are a common sight in airports, but could they soon be appearing in doctor’s rooms and hospitals?
There is a growing need for accurate yet inexpensive ways to screen for cancer at a population level, especially in low- and middle-income countries. Now, a new study suggests that combining canine olfaction and a particular type of statistical modelling can detect cancer from your breath with a high degree of accuracy.
“Because breath collection and canine detection are noninvasive, logistically simple, and low-cost, canine teams could plausibly support scalable population-level multicancer triage,” an Indian research team wrote in the Journal of Clinical Oncology.
The phase 2 assessor-masked case-control study was conducted in six Indian hospitals. Adults aged 18 years and older were eligible for the study. Cases required a biopsy- or fine-needle aspiration cytology-confirmed malignancy; controls could be healthy, have a nononcologic chronic condition, or a benign biopsy result. Individuals who were pregnant or who had a communicable respiratory illness were excluded.
There were 3275 participants enrolled in the study, with 1773 of these (1400 controls and 373 cases) used in the training cohort for the statistical models. The remaining 1502 participants (1219 cases and 283 controls) were used as the test cohort. All participants completed 10 minutes of tidal breathing while wearing a standard cotton surgical mask. Used masks were collected and stored until testing occurred.
The 1502 patients in the test cohort were predominantly female (62.7%) and had an average age of 49.7 years. Most had never smoked or chewed tobacco (91.0% and 80.4%, respectively) with roughly one in three patients (35.4%) reporting the presence of a chronic comorbidity (e.g., hypertension, diabetes, other heart diseases, epilepsy, or asthma). The largest number of patients in the test cohort had head and neck cancer (n = 88), followed by upper gastrointestinal cancers (n = 59) and breast cancer (n = 48).
Seven different dogs (six females; four beagles, one labrador, one labrador-indie mix, and one Dutch shepherd-Belgian Malinois mix) were used in the study. The dogs were exposed to the mask samples from the individuals in the training cohort, with the results of the canine olfactory detection and patient disease and demographic information combined into the Bayesian fusion framework. The fusion was then used to predict the likelihood of a given sample in the test cohort being positive for cancer.
When the test samples were analysed, the Bayesian fusion model performed well: 90.8% sensitivity (95% confidence interval 87.2-94.5) and 91.3% specificity (89.7-92.9), with a receiver operating characteristic area under the curve (ROC AUC) of 0.962 (0.952-0.969).
In terms of actual numbers, the model correctly classified 257 of the 283 patients with cancer and 1113 of the 1219 control patients – meaning there were 26 false negatives (failing to correctly detect a patient with cancer) and 106 false positives (incorrectly saying a control patient had cancer).
The fusion model performed at a similar level when different types and stages of cancer were considered, as well as across different sources of biological, behavioural, and operational variability.
The researchers concluded that the findings “support further evaluation of breath-based canine detection as a potential upstream triage approach within diagnostic pathways”.
The authors of an accompanying editorial were also impressed by the results but cautioned against putting the cart before the horse.
“At face value, these results are striking,” wrote Dr Shiran Shapira (PhD, a senior lecturer in human molecular genetics and biochemistry) and Professor Nadir Arber (a clinician and researcher with experience in the early detection of gastrointestinal cancers), both from the Division of Medicine at the Tel Aviv Sourasky University Hospital in Israel. “However, interpreting them requires careful consideration of both the biological plausibility and the methodological constraints inherent to studies of this type.”
It’s also critical to note that seven of the authors on the paper were employed by, had stock or other ownership interests in, had consulted or advised for, or had travel, accommodation and other expenses paid for by Dognosis, a research and development venture that uses dogs to detect disease. Dognosis also provided funding for the current study.
Dr Shapira and Professor Arber acknowledge the growing body of evidence suggesting that cancer affects cellular metabolism, which in turn may result in the production of characteristic volatile organic compound signals, and that dogs can be trained to recognise such signals. The pair also praise the current study to test whether dogs can smell cancer on your breath in a more structured and analytical manner.
“[The various] methodological safeguards strengthen the internal validity of the study and represent a meaningful step towards evaluating biological olfactory detection within the rigor expected of modern diagnostic research,” they wrote. “Nevertheless, whether such biologically driven detection systems can ultimately meet the scalability, reproducibility, and regulatory standards required for population-level cancer screening remains an open question.”
Related
The question of scalability remains a significant talking point, according to Dr Shapira and Professor Arber. “Dogs may smell cancer; but they cannot screen the world,” they state. “Unlike laboratory instruments, dogs cannot be manufactured, calibrated, and distributed at an industrial scale.”
“[Therefore,] the most transformative implication of canine detection studies may not be the deployment of dogs themselves, but rather the identification of the biochemical signals they detect. If the volatile signatures recognized by trained animals can be characterised chemically, they could inform the development of electronic noses, sensor arrays, or artificial intelligence–assisted breath analysers capable of replicating canine sensitivity in a scalable technological platform.”
Electronic noses and other wearable sensor arrays are already being used in research, although there are still challenges to be overcome before they can be deployed more broadly.
“The field of multicancer detection is evolving rapidly, with several emerging approaches. Most notably, circulating tumour DNA-based assays are already entering clinical evaluation. These technologies aim to identify tumour-derived molecular signals directly from blood samples,” said Dr Shapira and Professor Arber.
“Breath-based detection represents a conceptually different strategy, targeting systemic metabolic changes rather than tumour-derived DNA fragments. Whether such signals prove complementary to existing [multicancer early detection] technologies or ultimately superior/inferior in accuracy remains to be determined.
“Notably, breath-based approaches offer potential advantages in cost, accessibility, and patient acceptability, attributes that are particularly relevant in LMIC settings where the burden of cancer continues to rise.
“The future of cancer screening may therefore depend less on searching for tumours themselves and more on decoding the systemic biological signals they leave behind… By focusing on host response rather than tumour-derived signals alone, this strategy may enable detection of malignancies at earlier stages while remaining compatible with scalable clinical laboratory workflows.”
