1. Analysis of Exhaled Breath with Electronic Nose and Diagnosis of Lung Cancer by Support Vector Machine
Dr.med. Māris Bukovskis 1 2 3, Dr.biol. Gunta Strazda 1 2 3, Dr.med Uldis Kopeika 3 4, Dr.biol.Normunds Jurka 3, Dr. Ainis Pirtnieks 4, Ph.dr. Līga Balode 3, Dr. Jevgenija Aprinceva 2, Inara Kantane 5, Prof. Immanuels Taivans 1 2 3
1 Center of Lung Diseases, Pauls Stradins Clinical University Hospital,
2 Faculty of Medicine, University of Latvia,
3 Institute of Experimental and Clinical Medicine, University of Latvia,
4 Department of Thoracic Surgery, Pauls Stradins Clinical University Hospital
5 Faculty of Economics and Management, University of Latvia
Dear collegues, dear professors I would like to thank the organizers of the congress for this opportunity to present the results of our study.

2. Conflict of interests No conflict of interests
Study was sponsored by ERAF activity Project Nb. 2010/0303/2DP/ /10/APIA/ VIAA/043/
The study does not have a conflict of interests and is sponsored by the European Regional Developement Foundation and University of Latvia.

3. Lung cancer mortality and diagnostic methods
Lung cancer causes 1.3 million deaths annually, more than the next three most common cancers (colon, breast and prostate) combined
% of patients with stage I lung cancer survive for 5 years
For distant tumors the 5-year survival rate is only 3.5 %
Available diagnostic methods - nonsensitive, expensive or invasive
Lung cancer is the most common cancer worldwide, accounting for 1.3 million deaths annually more than next three most common cancers combined. Survival rates for lung cancer are generally lower than those for most cancers. 5-year survival rate for lung cancer is about 16%, compared to 65% for colon cancer, 89% for breast cancer, and over 99% for prostate cancer.
The diagnosis of early stage lung cancer is essential and substantially determines the 5-year life expectancy % of patients with stage I lung cancer survive for 5 years. 5-year survival of patients with stage III lung cancer is only %. For distant tumors (spread to other organs) the five-year survival rate is only 3.5 percent. However, only 15 percent of lung cancer cases are diagnosed at an early stage. Over half of people with lung cancer die within one year of being diagnosed.
The problem is that available diagnostic methods, i.e., sputum cytologic analyis and chest x-ray are nonsensitive, CT or PET – expensive, but FB and biopsy - invasive. They are not fit for patient screening.
Actually currently there is no good screening method of lung cancer risk groups.
World Health Organization. Cancer Fact Sheet 2009
American Cancer Society. Cancer Facts & Figures 2012

4. Lung cancer sniffer dogs
VOC’s in exhaled breath
Lung cancer sniffer dogs
CBC News Aug 17, 2011
One possible screening option of lung cancer is the analysis of exhaled air.
Studies have shown that there are hundreds of volatile organic compounds (VOCs) in human exhaled air. They are mainly blood born and therefore enable monitoring of different processes in the body.
Early studies in 1980s using GC-MS method demostrated difference in VOCs in patients with lung cancer and healthy controls.
Some studies have demonstrated that dogs are able to distinguish between breath samples from people with lung cancer from those with COPD with accuracy up to 71%.
Studies during past ten years have demonstrated possibility to discriminate lung cancer in exhaled air by principally new technology – electronic nose.
Gordon SM et al. Clin Chem 1985
Machado et al. AJRCCM 2005
Chen X et al. Cancer 2007

5. Functional principles of electronic nose
VOCs induce change of the sensor
volume and subsequently change of
electric resistance
A unique response curve combination,
containing the information to allow
discrimination of the different samples
Cyranose 320 e-
Electronic nose is an instrument made up of chemical sensors combined with a pattern recognition system. The measurement in EN is based on the differential electrical resistance response of multiple sensors that differs regarding VOC’s molecular pattern. If there has been no change in the composition of the air the volume and resistance of sensors does not change. When sensors are exposed to VOCs, the polymer matrix in Cyranose acts like a sponge and swells while absorbing the analyte and subsequently change the electric resistance.
Each sensor may give response to different VOCs and many sensors to one. A unique response curve combination is recorded, containing the information to allow discrimination of the different samples.
If we examine patient groups with different diseases, pattern recognition system develops typical average “smellprint” characteristic for patients with similar diagnosis.
VOCs
S S S S S S6

6. Objective
The aim of our study was to prove the potential of exhaled breath analysis and Support Vector Machine (SVM) to discriminate patients with:
1) lung cancer from healthy controls and other lung diseases;
2) lung cancer with or without COPD from patients with only COPD and healthy controls;
3) early stage lung cancer.
The aim of our study was to prove the potential of exhaled breath analysis and Support Vector Machine (SVM) to discriminate patients with: 1) lung cancer from healthy controls and other lung diseases, 2) lung cancer with or without COPD from patients with only COPD and healthy controls, 3) early stage lung cancer.

7. Dragonieri S et al. J Allergy Clin Immunol 2007
Methods
Sampling of exhaled air
Inspiration of VOC-filtered air by tidal breathing for 5 minutes, through T-shaped two-way non-rebreathing valve (Hans Rudolph Inc., Shawnee, USA)
Inhalation to total lung capacity and full exhalation with approximate flow rate 0.25 – 0.5 L/s into a polyethylene terephthalate plastic bag
Analysis by electronic nose device (Cyranose 320, Smith Detection, USA) within 5 minutes after breath sample collection
First crucial question is sampling of the exhaled air.
Initially, patients breathe tidally with clipped nose, through filtered T-shaped two-way non-rebreathing valve for 5 min. It is necessary to clean the exhaled breath from ambient air pollution, which can be substantial in hospital. Single inhalation is not sufficent to equilibrate the exhaled breath VOCs.
Second, inhalation to total lung capacity and full exhalation into polyethylene terephthalate plastic bag.
And, third, immediate analysis within 5 min. after sampling by EN device.
Dragonieri S et al. J Allergy Clin Immunol 2007

8. Methods Satistical analysis Support vector machine (SVM)
Continuous predictors: relative maximum (Rmax), area under curve (∑0-60”) and tg α0-60” for each curve of 32 sensors
Additional predictor factors: age, smoking status (smoker, non-smoker, ex-smoker), smoking history (pack-years) and ambient temperature tº C at the moment of measurement
Next, statistical analysis by pattern recognition system, in this case – support vector machine (SVM).
Relative maximum (Rmax), area under curve (AUC0-60” ) and tgα0-60” for each curve of 32 sensors were calculated and used as continuous predictors in SVM analysis. Age, smoking status, smoking history and ambient temperature were used as additional predictor factors.

9. Support Vector Machine
SVM is a method of supervised learning models with associated learning algorithms that analyze data and recognize patterns, used for classification analysis. SVM constructs a hyperplane or set of hyperplanes in a high- or infinite-dimensional space, which can be used for classification, regression, or other tasks. The SVM takes a set of input data and predicts, for each given input, which of two possible classes forms the output, making it a non-probabilistic binary linear classifier.

10. Results Morphologically confirmed lung cancer
166 patients with lung cancer, 91 patients with other diagnoses (COPD, pneumonia, TB, PATE, benign tumors etc.) and 79 healthy volunteers were included in this analysis.
Morphologically confirmed lung cancer
Other diseases: COPD, pneumonia, tbc, PATE, benign tumors etc.
Control – healthy volunteers, postinflammatory pneumofibrosis

11. Results Cancer vs No cancer
Parameters of 32 detectors Rmax, ∑0-60 un tg α0-60 Age, Pack-years and ambient tºC
Cross-validation 72.8% Class accuracy 79.1%
Classification summary (Support Vector Machine), Cancer vs No cancer, Training/Test sample 100% SVM: Classification type 1 (C=2.000), Kernel: Linear Number of support vectors = 219 (170 bounded) Include criteria: v20='GF'
Total
Correct
Incorrect
Correct (%)
Incorrect (%)
Cancer
165
144 21
87.3
12.7
No cancer
170
121 49
71.2
28.8
On this slide you can see classification summary of SVM of cancer group and non cancer group. Test group is the same as training group.
144 patients out of 165 cancer were classified correctly.
Sensitivity of the method is 87.3%, specificity 71.2%, PPV 74.6% and NPV 85.2%. Cross-validation value was 72.8%.
Cancer
No cancer
144 49
74.6
PPV 21
121
85.2
NPV
87.3
71.2
Sensitivity
Specificity

12. Results Cancer vs No cancer
Parameters of 32 detectors Rmax, ∑0-60 un tg α0-60 Age, Pack-years and ambient tºC
Cross-validation 69.7% Class accuracy 75.5%
Classification summary (Support Vector Machine), Cancer vs No cancer, Training sample 75% Test sample 25% SVM: Classification type 1 (C=2.000), Kernel: Linear Number of support vectors = 219 (170 bounded) Include criteria: v20='GF'
Total
Correct
Incorrect
Correct (%)
Incorrect (%)
Cancer 45 40 5
88.9
11.1
No cancer 39 26 13
66.7
33.3
Division of all study population into training 75% and test group 25% gives correct classification of 40 out of 45 cancer patients in the test group.
Sensitivity of the method is 88.9%, specificity 66.7%, PPV 75.5% and NPV 83.9%, with cross-validation value 69.7%.
Cancer
No cancer 40 13
75.5
PPV 5 26
83.9
NPV
88.9
66.7
Sensitivity
Specificity

13. Results Cancer vs Control
Parameters of 32 detectors Rmax, ∑0-60 un tg α0-60 Age, Pack-years and ambient tºC
Cross-validation 90.6% Class accuracy 93.1%
Classification summary (Support Vector Machine) Cancer vs Control Training/Test sample 100% SVM: Classification type 1 (C=5.000), Kernel: Linear Number of support vectors = 84 (39 bounded) Include criteria: v20='GF'
Total
Correct
Incorrect
Correct (%)
Incerrect (%)
Cancer
166
164 2
98.8
1.2
Control 79 64 15
81.0
19.0
Classification summary of SVM of cancer and control group shows high cross-validation accuracy 90.6%. 164 out of 166 cancer patients in training group were predicted correctly. We get high sensitivity value 98.8% nad NPV value 97.0%.
Cancer
Control
164 15
91.6
PPV 2 64
97.0
NPV
98.8
81.0
Sensitivity
Specificity

14. Results Cancer vs Control
Parameters of 32 detectors Rmax, ∑0-60 un tg α0-60 Age, Pack-years and ambient tºC
Cross-validation 89.7% Class accuracy 93.5%
Classification summary (Support Vector Machine) Cancer vs Control Training sample 75% Test sample 25% SVM: Classification type 1 (C=5.000), Kernel: Linear Number of support vectors = 84 (39 saistīti) Include criteria: v20='GF'
Total
Correct
Incorrect
Correct (%)
Incorrect (%)
Cancer 45 44 1
97.8
2.2
Control 16 11 5
68.8
31.2
Only case out of 45 is classified incorrectly in the test group, if we divide this population into training 75% and test group 25% .
Sensitivity is 97.8%, relative low specificity 68.8%, PPV 89.8% and NPV 91.7%. Cross-validation value 89.7%.
Cancer
Control 44 5
89.8
PPV 1 11
91.7
NPV
97.8
68.8
Sensitivity
Specificity

15. Results Cancer vs Cancer + COPD vs COPD vs Control
Parameters of 32 detectors Rmax, ∑0-60 un tg α0-60 Age, Pack-years and Ambient tºC
Cross-validation 71.1% Class accuracy 77.4%
Classification summary (Support Vector Machine), Cancer vs Cancer+COPD vs COPD vs Control, SVM: Classification type 1 (C=2.000), Kernel: Linear Number of support vectors = 152 (43 bounded) Include criteria: v20='GF'
Total
Correct
Incorrect
Correct (%)
Incorrect (%)
Cancer 63 36 27
57.1
42.9
Cancer + COPD 79
100.0
0.0
COPD 15 5 10
33.3
66.7
Control 78 62 16
79.5
20.5
Both lung cancer and COPD are diseases associated with cigarette smoking and coexist in many patients. There have been previous attempts to differentiate patients with lung cancer from those with COPD in clinical studies of exhaled breath analysis by electronic nose, yet many patients have both conditions at the same time, and results are ambiguous.
Exhaled breath of morphologically verified lung cancer patients (cancer group), lung cancer patients with concomitant COPD (cancer and COPD group), COPD patients without verified lung cancer (COPD group) and healthy volunteers (control group) was examined.
As can be seen from this table, high prognostic accuracy is achieved only in cancer + COPD group: 79 out of 79 patients are predicted accurately. In other patient groups rate of correct prognoses looks quite poor.

16. Results Cancer vs Cancer + COPD vs COPD vs Control
Parameters of 32 detectors Rmax, ∑0-60 un tg α0-60 Age, Pack-years and Ambient tºC
Cross-validation 71.1% Class accuracy 77.4%
Classification matrix (Support Vector Machine), Cancer vs Cancer+COPD vs COPD vs Control, Training/Test sample 100% SVM: Classification type1 (C=2.000), Kernel: Linear, Number of support vectors = 152 (43 bounded) Prognosis (rows) x Diagnosis (columns)
Cancer
Cancer + COPD
COPD
Control 36 26 1
79 (!) 9 5 7 62
But if we analyze classification matrix, it becomes obvious that only one patient with lung cancer is classified incorrectly.
The main classification error in cancer group occurred between cancer and cancer+COPD group.

17. Results
Parameters of 32 detectors Rmax, ∑0-60 un tg α0-60 Age, Pack-years and Ambient tºC
Patients with post-obstructive pneumonia in cancer group and bacterial,
TB and infarct pneumonia in no cancer group were excluded from analysis
Classification matrix (Support Vector Machine), Stage 1-2, 3 and 4 Training/Test group 100% SVM: Classification type 1 (C=1.000), Kernel: Linear, Number of support vectors = 184 (73 bounded) Prognosis (rows) x Diagnosis (columns)
No cancer
Stage 1-2
Stage 3
Stage 4
100 7 2 11 3 25 1 9 40 27
The diagnosis of early stage lung cancer is essential and substantially determines the 5-year life expectancy. The aim of this study was to prove the potential of exhaled breath analysis to discriminate an early stage lung cancer.
72.5% (29 out of 40) patients with stage 1-2 lung cancer and 82.3% (106 out of 135) patients with any stage of lung cancer were predicted correctly as the cancer group.
The exhaled breath analysis by electronic nose using support vector pattern recognition method is able to discriminate satisfactorily an early stage lung cancer in healthy subjects and patients with different lung diseases, if Patients with post-obstructive pneumonia in cancer group and bacterial, TB and infarct pneumonia in no cancer group were excluded from analysis.

18. Conclusions
Exhaled breath analysis by electronic nose using support vector pattern recognition method is able to discriminate:
Lung cancer from healthy subjects and patients with different lung diseases
An early stage lung cancer from healthy subjects and patients with different lung diseases
Some discrimination pattern between lung cancer, patients with lung cancer and COPD, COPD and control, even in patients with combined diseases
Exhaled breath analysis by electronic nose using support vector pattern recognition method is able to discriminate: 1) lung cancer from healthy subjects and patients with different lung diseases, 2) an early stage lung cancer from healthy subjects and patients with different lung diseases, 3) some discrimination pattern between lung cancer, patients with lung cancer and COPD, COPD and control, even in patients with combined diseases.

19. Acknowledgements To my colleagues and our team Prof. Immanuels Taivans
Dr.biol. Gunta Strazda
Dr. Ainis Pirtnieks
Dr.med. Uldis Kopeika
Dr.biol. Normunds Jurka
Ph.dr. Liga Balode
Doctoral student Agnese Kislina
Mrs. Inara Kantane
Finally I would like to thank my colleagues who took part in this work.

20. Thank You for Your Attention!
And the audience for your attention...!
How to sniff out the disease?