- Email: [email protected]
Distinguishing idiopathic Parkinson's disease from other parkinsonian syndromes by breath test
Parkinsonism and Related Disorders 21 (2015) 150e153
Contents lists available at ScienceDirect
Parkinsonism and Related Disorders journal homepage: www.elsevier.com/locate/parkreldis
Short communication
Distinguishing idiopathic Parkinson's disease from other parkinsonian syndromes by breath test M.K. Nakhleh a, 1, S. Badarny b, **, 1, R. Winer b, R. Jeries a, J. Finberg c, H. Haick a, * a
Department of Chemical Engineering and Russell Berrie Nanotechnology Institute, Technion - Israel Institute of Technology, Haifa 3200003, Israel Movement Disorders Clinic, Department of Neurology, Carmel Medical Center, and Faculty of Medicine, Technion e Israel Institute of Technology, Haifa 3200003, Israel c Department of Molecular Pharmacology, Bruce Rappaport Faculty of Medicine, Technion Israel Institute of Technology, Haifa 31096, Israel b
a r t i c l e i n f o
a b s t r a c t
Article history: Received 22 July 2014 Received in revised form 12 November 2014 Accepted 27 November 2014
Introduction: Diagnosis of different parkinsonian syndromes is linked with high misdiagnosis rates and various confounding factors. This is particularly problematic in its early stages. With this in mind, the current pilot study aimed to distinguish between Idiopathic Parkinson's Disease (iPD), other Parkinsonian syndromes (non-iPD) and healthy subjects, by a breath test that analyzes the exhaled volatile organic compounds using a highly sensitive nanoarray. Methods: Breath samples of 44 iPD, 16 non-iPD patients and 37 healthy controls were collected. The samples were passed over a nanoarray and the resulting electrical signals were analyzed with discriminant factor analysis as well as by a K-fold cross-validation method, to test the accuracy of the model. Results: Comparison of non-iPD with iPD states yielded 88% sensitivity, 88% accuracy, and 88% Receiver Operating Characteristic area under the curve in the training set samples with known identity. The validation set of this comparison scored 81% sensitivity and accuracy and 92% negative predictive value. Comparison between atypical parkinsonism states and healthy subjects scored 94% sensitivity and 85% accuracy in the training set samples with known identity. The validation set of this comparison scored 81% sensitivity and 78% accuracy. The obtained results were not affected by L-Dopa or MAO-B inhibitor treatment. Conclusions: Exhaled breath analysis with nanoarray is a promising approach for a non-invasive, inexpensive, and portable technique for differentiation between different Parkinsonian states. A larger cohort is required in order to establish the clinical usefulness of the method. © 2014 Elsevier Ltd. All rights reserved.
Keywords: Parkinson's disease/parkinsonism Breath test Nano-array Volatile organic compounds
1. Introduction Historically, the discrimination between Idiopathic Parkinson's Disease (iPD) and other parkinsonian disorders (non-iPD) has relied mainly on clinical features and physician evaluation [1e3].
* Corresponding author. Department of Chemical Engineering and Russell Berrie Nanotechnology Institute, Technion - Israel Institute of Technology, Haifa 3200003, Israel. ** Corresponding author. Movement Disorders Clinic, Department of Neurology, Carmel Medical Center, Haifa 3436212, Israel. Tel.: þ972 4 8250390; fax: þ972 4 8250693. E-mail addresses: [email protected] (S. Badarny), [email protected] (H. Haick). 1 Authors have equal contribution to the manuscript. http://dx.doi.org/10.1016/j.parkreldis.2014.11.023 1353-8020/© 2014 Elsevier Ltd. All rights reserved.
However, re-examination of such patients indicates misdiagnosis rates of about 20% in the parkinsonian population [1,3], while pathological brain tests of iPD-diagnosed patients can lead to ca. 35% misclassification rates [1,3]. A variety of imaging tools such as PET, SPECT, MRI, Diffusion Tensor MRI, and Transcranial Sonography (TCS) exist which can be used to differentiate between different parkinsonian conditions [4]. These techniques can detect abnormalities in the basal ganglia even before clinical signs appear, but do not reliably distinguish between different neurodegenerative parkinsonian syndromes in all cases, and the detected abnormalities do not always correlate well with disease progression [2,5]. There is much effort being directed towards development of biomarkers for iPD, but no suitable biomarker is yet available for general use [2]. An additional practical problem is that many of the more sophisticated techniques are not generally available to
M.K. Nakhleh et al. / Parkinsonism and Related Disorders 21 (2015) 150e153
practitioners because of the requirement for special equipment or expensive radioactive materials. Measurements of alpha-synuclein levels in cerebro-spinal fluid (CSF) score high accuracy for iPD detection, but not in blood or saliva samples [2]. Here, we report on a preliminary clinical study that aims to examine the possibility of discriminating iPD from other parkinsonian (non-iPD) states by exhaled breath analysis. The rationale of this study is based on earlier evidence that neurological and other pathological processes might result in an altered spectrum of volatile organic compounds that are excreted via exhaled breath [6]. In a pre-clinical study, for example, Tisch et al. [7] have reported an association between asymptomatic nigrostriatal dopaminergic lesion, in an animal model, and volatile biomarkers in exhaled breath samples. Another source of this rationale is based on our previous proof-of-concept clinical study in which we were able to discriminate between Parkinson's and Alzheimer diseases by analysis of exhaled breath samples [8]. While breath analysis can be achieved using mass spectrometry techniques, we focus our current paper on the use of a novel system, consisting of an array of monolayer capped nanoparticle sensors, because of its high potential to serve as a point-of-care diagnostic tool [9]. When conjugated with pattern recognition methods, the nanoarray can be
151
trained to identify a standardized group of volatiles, and was inspired by sensitivity of the mammalian sense of smell [9]. 2. Methods 2.1. Population and breath sampling 44 iPD, 16 non-iPD (including 4 Drug induced, 3 Progressive Supranuclear Palsy (PSP), 2 Multiple System Atrophy (MSA), 1 Cortical Basal Ganglionic Degeneration (CBGD), 2 Diffuse Lewy Body Disease (DLBD) and 4 patients with combined subtypes), and 37 healthy subjects were enrolled to the study. Breath samples were obtained by the same technician, at the same examination room and environment, in a controlled manner, in which, the inhaled ambient air is filtered and the dead space volume is excluded. As a precaution, the participants were directed not to consume food (besides their medication), alcohol or tobacco, and not to perform physical exercise overnight prior to sampling. On each sampling day, subjects from at least two of the study groups were enrolled in a randomized manner. The diagnosis of each of the patients was determined according to clinical examination by an experienced movement disorders specialist, Dr. Samih Badarny. Periodic reexamination of the cohort has been carried out for a period of three years, in order to substantiate the clinical diagnosis. Study groups were age and sex matched, and clinical features of the recruited volunteers are summarized in Supplementary Table S1. Two samples of end-tidal exhaled breath were collected from each subject and volatile compounds contained in the breath were adsorbed on two-bed ORBO™ 420 Tenax TA sorption tubes (SigmaeAldrich, St. Louis, MO, USA) which were sealed and stored at 4 C until analysis. Samples were analyzed during 5 consecutive days after the collection phase was over. All participants signed an informed consent
Fig. 1. (A) K-Fold cross-validation method was employed to calculate the prediction accuracy of DFA models. The classifier was computed 4 times. In each time, we have used 33 iPD and 12 non-iPD samples as training set, while the remaining 15 samples were left out as validation set; each sample was considered once as part of the training set and once as part of the validation set. (B) DFA-model (i.e. graphical presentation of the first canonical score) discriminating between parkinsonism and iPD. (C) Receiver operating characteristic (ROC) curves for the data of the training sets discriminated between parkinsonism and iPD. (D) DFA-model (i.e. graphical presentation of the first canonical score) discriminating between parkinsonism and healthy subjects. (E) Receiver operating characteristic (ROC) curves for the data of the training sets discriminated between parkinsonism and healthy subjects. (F) Receiver operating characteristic (ROC) curve for the data discriminating between L-Dopa (black squares)-treated or MAO inhibitor-treated (blue triangle) and nontreated parkinsonian subjects. The boxes represent the 95% confidence intervals, corresponding to 1.96 SEM. AUC ¼ Area Under ROC Curve. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
90% 59%
form. The study was approved by the local IRB committee of the Carmel Hospital, Haifa, Israel. For comprehensive details regarding diagnostic criteria and breath sampling technique see Supplementary Material (Sections 1.1. and 1.2.). 2.2. Nanoarray analysis and statistical analysis Compounds trapped on the Tenax tubes were thermally released (250 C/ 10 min) and then injected into an exposure cell containing an array of sensors that are based on chemically-modified gold nanoparticles or single walled carbon nanotubes [9]. The interaction between the films of the modified particles and the volatile organic compounds results in a time dependent and reversible change in resistance of the sensors [9]. The diversity of organic coating layers results in a variety of signals that were recorded and then analyzed using discriminate factor analysis (DFA) [10], in order to establish a binary classifier. DFA is a supervised linear method in which sensor responses are inserted as primary input, and a linear combination of the variables is calculated to produce new orthogonal axes (canonical variables) while minimizing the variance within each class and maximizing the variance between the classes, thus choosing the sensors most contributive to the predictive model. In order to prevent over-fitting of the data, we used a ratio of features to samples lower than 1:6. K-Fold cross-validation method was employed to calculate the prediction accuracy of DFA models. The classifier was computed 4 times. In each time, we have used 33 iPD and 12 non-iPD samples as training set, while the remaining 15 samples were left out as validation set; each sample was considered once as part of the training set and once as part of the validation set, then, we averaged the results of all 4 training and validation sets (Fig. 1). Statistical calculations were performed using SAS JMP, Version 10.0. More detailed information on the nanoarray and analysis can be found elsewhere (See Supplementary Materials Sections 1.3. and 1.4.).
3. Results
81%
f
e
c
d
Sensitivity ¼ (True Positive/True positive þ False negative). Specificity ¼ (True Negative/True Negative þ False Positive). Accuracy ¼ (True Positive þ True Negative/n). Positive Predictive Value (PPV ¼ True positive/True positive þ False Positive). Negative Predictive Value (NPV ¼ True Negative/True Negative þ False Negative). Average of all 4 Training/Validation sets. a
2
b
81% 94%
85%
85%
68%
97%
13
3
30
7
76%
78%
92% 62% 81% 81% 81% 8 36 3 13 95% 74% 88% 88% 88% 88%
iPD n ¼ 44 Healthy n ¼ 37 non-iPD n ¼ 16 Non-iPD n ¼ 16 1
Training set Population
Table 1 Statistical analysis of training and validation sets of K-fold analysis.
Validation set
M.K. Nakhleh et al. / Parkinsonism and Related Disorders 21 (2015) 150e153
Classifier Target group Control group Sensitivitya,f Specificityb,f Accuracyc,f ROC AUC PPVd,f NPVe,f True positive False negative True negative False positive Sensitivitya,f Specificityb,f Accuracyc,f PPVd,f NPVe,f
152
Two DFA classifiers were established and statistical analysis was performed both for the training and validation sets (Fig. 1A). The first classifier discriminated between iPD and non-iPD subjects with 88% sensitivity and 88% accuracy in the training set samples of known identity, when targeting non-iPD. The validation set of this comparison scored 81% sensitivity and 81% accuracy. Moreover, the model scored a Negative Predictive Value (NPV) of 92% (Fig. 1 and Table 1). The area under Receiver Operating Characteristic (ROC) was 88% (Fig. 1C). The second classifier discriminated between noniPD patients and healthy subjects with 94% sensitivity, 81% specificity, and 97% NPV in the training set samples of known identity (Fig. 1D and Table 1). In the validation set this comparison scored 81% sensitivity, 76% specificity, and 90% NPV. The area under ROC was 85% (Fig. 1E). The confounding effect of L-dopa or MAO-B inhibitor treatment was tested by applying the first classifier on all patients, trying to distinguish between de-novo and L-dopa or MAOB inhibitor-treated parkinsonian patients. The results of this test showed only a random classification of the samples with ROC AUC of 56% and 59%, respectively (Fig. 1F). It is noteworthy to point out that the signals obtained when analyzing parkinsonian samples were more variant and less centralized in comparison to iPD or healthy control samples. This variance was also expressed in the relatively low PPV (62%) in the classifier validation, and could be explained by the heterogeneity of this group, which contains several sub-types: Drug Induced Parkinsonism (25%), Progressive Supranuclear Palsy (19%), Multiple System Atrophy (13%), CorticoBasal Ganglionic Degeneration (6%), Diffuse Lewy Body Disease (13%) and combined sub-types (25%) (Supplementary Table S1). On the other hand, the iPD group was much more homogenous, enabling better detection ability of the classifier (NPV ¼ 92%). 4. Discussion In this study we report on a preliminary result indicating that discrimination between iPD and non-iPD parkinsonism could be achieved by breath analysis using an array of nanomaterials based gas sensors with an accuracy of 81%. However, a potentially obvious limitation of the current study is that the diagnosis was based upon the specialist examination only, which could lack accuracy, instead
M.K. Nakhleh et al. / Parkinsonism and Related Disorders 21 (2015) 150e153
of the gold standard brain autopsy or by a combination of imaging techniques. On the other hand, periodic re-examination of the cohort has been carried out for 3 years, in order to monitor disease progression, development of clinical symptoms and the response to treatment, which together are used as confirmation of the initial diagnosis. The preliminary results presented here indicate a powerful ability of breath analysis to distinguish between non-iPD, iPD and healthy subjects. These results are comparable to transcranial sonography, which can discriminate between iPD and PSP or MSA, with 91% sensitivity and 82% specificity. However, the nanoarray has a great advantage of being more accurate in distinguishing between iPD and other parkinsonian syndromes [11]. The nanoarray has lower sensitivity (88%) than that of the olfactory disturbance among iPD cases (96%). Nevertheless, the nanoarray has higher specificity (81%) than that of the olfactory disturbance (63%) when the discrimination between iPD and MSA is targeted [12]. The alteration in the response of the nanoarray to exhaled air of parkinsonism compared to normal subjects is due to the overall changes in the content of exhaled breath volatile biomarkers. This observation was supported via parallel analysis with gaschromatography/mass spectrometry (GCeMS), which revealed tentatively 4 volatile organic compounds for the discrimination between iPD and non-iPD states (p-value < 0.05): 3methylhexane (11 ppb vs. 10 ppb), 2-pentanone (19 ppb vs. 4 ppb), benzaldehyde (31 ppb vs. 27 ppb); acetophenone (13 vs. 11 ppb), and 2,6 dimethylnonane. The same GCeMS analysis revealed three volatile biomarkers for the discrimination between parkinsonism and healthy states (p-value < 0.05): acetophenone (11 ppb vs. 13 pbb); 3-methyl hexane (10 ppb vs. 11 ppb); and 2,6-dimethylnonane. These significant alterations could be due to the effect of one or more of the following processes: oxidative stress, cytochrome p450, hepatic enzymes, carbohydrate-metabolism (glycolysis or gluconeogenesis pathways), lipid metabolism and others associated with the different neurological pathologies [6]. The origin and pathway of the reported compounds are now under extensive studies. Previously, it was evidenced that the nanoarray system is not influenced by potential confounding factors, such as age, gender, smoking and residence location [13,14]. In the current study, the results show no confounding influence of the L-dopa and/or MAO-B treatment on the performance of the applied nanoarray. This is because the nanoarray contains cross-reactive sensors that respond to most/all compounds in a mixture [9], rather than towards a specific biomarker. Therefore, any change in a single volatile marker of a given confounding factor would have negligible effect on the sensing signal, compared to the signal generated by the overall volatile compounds in the breath sample. In summary, we show that exhaled breath analysis with nanoarray is a promising non-invasive technique for distinguishing between different Parkinsonian states. Further development of the technique can produce a portable and inexpensive device which could be available at the level of individual clinics, however a larger clinical study is needed to validate the current results. Study funding This study was supported by the Kamin Fund, Israel Ministry of Trade and Industry (grant no. 47247).
153
Authors contribution Dr. Morad Nakhleh e Statistical analysis, interpretation of data and manuscript drafting. Final approval of the version to be submitted. Dr. Samih Badarny e Study design and supervision, and manuscript revision. Final approval of the version to be submitted. Dr. Raz Winer e Acquisition of clinical samples and data, revision of the manuscript for important intellectual content. Mrs. Raneen Jeries e Acquisition of clinical samples and data interpretation. Prof. John Finberg e Data interpretation and manuscript revision. Final approval of the version to be submitted. Prof. Hossam Haick e Study design and supervision, interpretation of data and manuscript revision. Final approval of the version to be submitted.
Financial disclosure All co-authors report no financial disclosures.
Appendix A. Supplementary data Supplementary data related to this article can be found at http:// dx.doi.org/10.1016/j.parkreldis.2014.11.023.
References [1] Hughes AJ, Daniel SE, Ben-Shlomo Y, Lees AJ. The accuracy of diagnosis of parkinsonian syndromes in a specialist movement disorder service. Brain 2002;125:861e70. [2] Schapira H. Recent developments in biomarkers in Parkinson disease. Curr Opin Neurol 2013;26:395e400. [3] Schrag A, Ben-Shlomo Y, Quinn N. How valid is the clinical diagnosis of Parkinson's disease in the community? J Neurol Neurosurg Psychiatry 2002;73: 529e34. [4] Berti V, Pupi A, Mosconi L. PET/CT in diagnosis of movement disorders. Ann N Y Acad Sci 2011. [5] Vogt T, Kramer K, Gartenschlaeger M, Schreckenberger M. Estimation of further disease progression of Parkinson's disease by dopamin transporter scan vs clinical rating. Parkinsonism Relat Disord 2011;17:459e63. [6] Haick H, Broza YY, Mochalski P, Ruzsanyi V, Amann A. Assessment, origin, and implementation of breath volatile cancer markers. Chem Soc Rev 2014;43: 1423e49. [7] Tisch U, Aluf Y, Ionescu R, Nakhleh M, Bassal R, Axelrod N, et al. Detection of asymptomatic nigrostriatal dopaminergic lesion in rats by exhaled air analysis using carbon nanotube sensors. ACS Chem Neurosci 2011;3:161e6. [8] Tisch U, Schlesinger I, Ionescu R, Nassar M, Axelrod N, Robertman D, et al. Detection of Alzheimer's and Parkinson's disease from exhaled breath using nanomaterial-based sensors. Nanomedicine (Future Medicine) 2013;8:43e56. [9] Broza YY, Haick H. Nanomaterial-based sensors for detection of disease by volatile organic compounds. Nanomedicine (Future Medicine) 2013;8: 785e806. [10] Breteton R. Chemometrics, application of mathematics and statistics to laboratory systems. J Chemometr 1990;6:228. [11] Berg D, Godau J, Walter U. Transcranial sonography in movement disorders. Lancet Neurol 2008;7:1044e55. [12] Haehner A, Hummel T, Reichmann H. Olfactory loss in Parkinson's disease. Parkinsons Dis 2011;450939:21. [13] Cohen-Kaminsky S, Nakhleh M, Perros F, Montani D, Girerd B, Garcia G, et al. A proof of concept for the detection and classification of pulmonary arterial hypertension through breath analysis with a sensor array. Am J Respir Crit Care Med 2013;188:756e9. [14] Nakhleh MK, Jeries R, Gharra A, Binder A, Broza YY, Pascoe M, et al. Detecting active pulmonary tuberculosis with a breath test using nanomaterial-based sensors. Eur Respir J 2014;43:1522e5.
Contents lists available at ScienceDirect
Parkinsonism and Related Disorders journal homepage: www.elsevier.com/locate/parkreldis
Short communication
Distinguishing idiopathic Parkinson's disease from other parkinsonian syndromes by breath test M.K. Nakhleh a, 1, S. Badarny b, **, 1, R. Winer b, R. Jeries a, J. Finberg c, H. Haick a, * a
Department of Chemical Engineering and Russell Berrie Nanotechnology Institute, Technion - Israel Institute of Technology, Haifa 3200003, Israel Movement Disorders Clinic, Department of Neurology, Carmel Medical Center, and Faculty of Medicine, Technion e Israel Institute of Technology, Haifa 3200003, Israel c Department of Molecular Pharmacology, Bruce Rappaport Faculty of Medicine, Technion Israel Institute of Technology, Haifa 31096, Israel b
a r t i c l e i n f o
a b s t r a c t
Article history: Received 22 July 2014 Received in revised form 12 November 2014 Accepted 27 November 2014
Introduction: Diagnosis of different parkinsonian syndromes is linked with high misdiagnosis rates and various confounding factors. This is particularly problematic in its early stages. With this in mind, the current pilot study aimed to distinguish between Idiopathic Parkinson's Disease (iPD), other Parkinsonian syndromes (non-iPD) and healthy subjects, by a breath test that analyzes the exhaled volatile organic compounds using a highly sensitive nanoarray. Methods: Breath samples of 44 iPD, 16 non-iPD patients and 37 healthy controls were collected. The samples were passed over a nanoarray and the resulting electrical signals were analyzed with discriminant factor analysis as well as by a K-fold cross-validation method, to test the accuracy of the model. Results: Comparison of non-iPD with iPD states yielded 88% sensitivity, 88% accuracy, and 88% Receiver Operating Characteristic area under the curve in the training set samples with known identity. The validation set of this comparison scored 81% sensitivity and accuracy and 92% negative predictive value. Comparison between atypical parkinsonism states and healthy subjects scored 94% sensitivity and 85% accuracy in the training set samples with known identity. The validation set of this comparison scored 81% sensitivity and 78% accuracy. The obtained results were not affected by L-Dopa or MAO-B inhibitor treatment. Conclusions: Exhaled breath analysis with nanoarray is a promising approach for a non-invasive, inexpensive, and portable technique for differentiation between different Parkinsonian states. A larger cohort is required in order to establish the clinical usefulness of the method. © 2014 Elsevier Ltd. All rights reserved.
Keywords: Parkinson's disease/parkinsonism Breath test Nano-array Volatile organic compounds
1. Introduction Historically, the discrimination between Idiopathic Parkinson's Disease (iPD) and other parkinsonian disorders (non-iPD) has relied mainly on clinical features and physician evaluation [1e3].
* Corresponding author. Department of Chemical Engineering and Russell Berrie Nanotechnology Institute, Technion - Israel Institute of Technology, Haifa 3200003, Israel. ** Corresponding author. Movement Disorders Clinic, Department of Neurology, Carmel Medical Center, Haifa 3436212, Israel. Tel.: þ972 4 8250390; fax: þ972 4 8250693. E-mail addresses: [email protected] (S. Badarny), [email protected] (H. Haick). 1 Authors have equal contribution to the manuscript. http://dx.doi.org/10.1016/j.parkreldis.2014.11.023 1353-8020/© 2014 Elsevier Ltd. All rights reserved.
However, re-examination of such patients indicates misdiagnosis rates of about 20% in the parkinsonian population [1,3], while pathological brain tests of iPD-diagnosed patients can lead to ca. 35% misclassification rates [1,3]. A variety of imaging tools such as PET, SPECT, MRI, Diffusion Tensor MRI, and Transcranial Sonography (TCS) exist which can be used to differentiate between different parkinsonian conditions [4]. These techniques can detect abnormalities in the basal ganglia even before clinical signs appear, but do not reliably distinguish between different neurodegenerative parkinsonian syndromes in all cases, and the detected abnormalities do not always correlate well with disease progression [2,5]. There is much effort being directed towards development of biomarkers for iPD, but no suitable biomarker is yet available for general use [2]. An additional practical problem is that many of the more sophisticated techniques are not generally available to
M.K. Nakhleh et al. / Parkinsonism and Related Disorders 21 (2015) 150e153
practitioners because of the requirement for special equipment or expensive radioactive materials. Measurements of alpha-synuclein levels in cerebro-spinal fluid (CSF) score high accuracy for iPD detection, but not in blood or saliva samples [2]. Here, we report on a preliminary clinical study that aims to examine the possibility of discriminating iPD from other parkinsonian (non-iPD) states by exhaled breath analysis. The rationale of this study is based on earlier evidence that neurological and other pathological processes might result in an altered spectrum of volatile organic compounds that are excreted via exhaled breath [6]. In a pre-clinical study, for example, Tisch et al. [7] have reported an association between asymptomatic nigrostriatal dopaminergic lesion, in an animal model, and volatile biomarkers in exhaled breath samples. Another source of this rationale is based on our previous proof-of-concept clinical study in which we were able to discriminate between Parkinson's and Alzheimer diseases by analysis of exhaled breath samples [8]. While breath analysis can be achieved using mass spectrometry techniques, we focus our current paper on the use of a novel system, consisting of an array of monolayer capped nanoparticle sensors, because of its high potential to serve as a point-of-care diagnostic tool [9]. When conjugated with pattern recognition methods, the nanoarray can be
151
trained to identify a standardized group of volatiles, and was inspired by sensitivity of the mammalian sense of smell [9]. 2. Methods 2.1. Population and breath sampling 44 iPD, 16 non-iPD (including 4 Drug induced, 3 Progressive Supranuclear Palsy (PSP), 2 Multiple System Atrophy (MSA), 1 Cortical Basal Ganglionic Degeneration (CBGD), 2 Diffuse Lewy Body Disease (DLBD) and 4 patients with combined subtypes), and 37 healthy subjects were enrolled to the study. Breath samples were obtained by the same technician, at the same examination room and environment, in a controlled manner, in which, the inhaled ambient air is filtered and the dead space volume is excluded. As a precaution, the participants were directed not to consume food (besides their medication), alcohol or tobacco, and not to perform physical exercise overnight prior to sampling. On each sampling day, subjects from at least two of the study groups were enrolled in a randomized manner. The diagnosis of each of the patients was determined according to clinical examination by an experienced movement disorders specialist, Dr. Samih Badarny. Periodic reexamination of the cohort has been carried out for a period of three years, in order to substantiate the clinical diagnosis. Study groups were age and sex matched, and clinical features of the recruited volunteers are summarized in Supplementary Table S1. Two samples of end-tidal exhaled breath were collected from each subject and volatile compounds contained in the breath were adsorbed on two-bed ORBO™ 420 Tenax TA sorption tubes (SigmaeAldrich, St. Louis, MO, USA) which were sealed and stored at 4 C until analysis. Samples were analyzed during 5 consecutive days after the collection phase was over. All participants signed an informed consent
Fig. 1. (A) K-Fold cross-validation method was employed to calculate the prediction accuracy of DFA models. The classifier was computed 4 times. In each time, we have used 33 iPD and 12 non-iPD samples as training set, while the remaining 15 samples were left out as validation set; each sample was considered once as part of the training set and once as part of the validation set. (B) DFA-model (i.e. graphical presentation of the first canonical score) discriminating between parkinsonism and iPD. (C) Receiver operating characteristic (ROC) curves for the data of the training sets discriminated between parkinsonism and iPD. (D) DFA-model (i.e. graphical presentation of the first canonical score) discriminating between parkinsonism and healthy subjects. (E) Receiver operating characteristic (ROC) curves for the data of the training sets discriminated between parkinsonism and healthy subjects. (F) Receiver operating characteristic (ROC) curve for the data discriminating between L-Dopa (black squares)-treated or MAO inhibitor-treated (blue triangle) and nontreated parkinsonian subjects. The boxes represent the 95% confidence intervals, corresponding to 1.96 SEM. AUC ¼ Area Under ROC Curve. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
90% 59%
form. The study was approved by the local IRB committee of the Carmel Hospital, Haifa, Israel. For comprehensive details regarding diagnostic criteria and breath sampling technique see Supplementary Material (Sections 1.1. and 1.2.). 2.2. Nanoarray analysis and statistical analysis Compounds trapped on the Tenax tubes were thermally released (250 C/ 10 min) and then injected into an exposure cell containing an array of sensors that are based on chemically-modified gold nanoparticles or single walled carbon nanotubes [9]. The interaction between the films of the modified particles and the volatile organic compounds results in a time dependent and reversible change in resistance of the sensors [9]. The diversity of organic coating layers results in a variety of signals that were recorded and then analyzed using discriminate factor analysis (DFA) [10], in order to establish a binary classifier. DFA is a supervised linear method in which sensor responses are inserted as primary input, and a linear combination of the variables is calculated to produce new orthogonal axes (canonical variables) while minimizing the variance within each class and maximizing the variance between the classes, thus choosing the sensors most contributive to the predictive model. In order to prevent over-fitting of the data, we used a ratio of features to samples lower than 1:6. K-Fold cross-validation method was employed to calculate the prediction accuracy of DFA models. The classifier was computed 4 times. In each time, we have used 33 iPD and 12 non-iPD samples as training set, while the remaining 15 samples were left out as validation set; each sample was considered once as part of the training set and once as part of the validation set, then, we averaged the results of all 4 training and validation sets (Fig. 1). Statistical calculations were performed using SAS JMP, Version 10.0. More detailed information on the nanoarray and analysis can be found elsewhere (See Supplementary Materials Sections 1.3. and 1.4.).
3. Results
81%
f
e
c
d
Sensitivity ¼ (True Positive/True positive þ False negative). Specificity ¼ (True Negative/True Negative þ False Positive). Accuracy ¼ (True Positive þ True Negative/n). Positive Predictive Value (PPV ¼ True positive/True positive þ False Positive). Negative Predictive Value (NPV ¼ True Negative/True Negative þ False Negative). Average of all 4 Training/Validation sets. a
2
b
81% 94%
85%
85%
68%
97%
13
3
30
7
76%
78%
92% 62% 81% 81% 81% 8 36 3 13 95% 74% 88% 88% 88% 88%
iPD n ¼ 44 Healthy n ¼ 37 non-iPD n ¼ 16 Non-iPD n ¼ 16 1
Training set Population
Table 1 Statistical analysis of training and validation sets of K-fold analysis.
Validation set
M.K. Nakhleh et al. / Parkinsonism and Related Disorders 21 (2015) 150e153
Classifier Target group Control group Sensitivitya,f Specificityb,f Accuracyc,f ROC AUC PPVd,f NPVe,f True positive False negative True negative False positive Sensitivitya,f Specificityb,f Accuracyc,f PPVd,f NPVe,f
152
Two DFA classifiers were established and statistical analysis was performed both for the training and validation sets (Fig. 1A). The first classifier discriminated between iPD and non-iPD subjects with 88% sensitivity and 88% accuracy in the training set samples of known identity, when targeting non-iPD. The validation set of this comparison scored 81% sensitivity and 81% accuracy. Moreover, the model scored a Negative Predictive Value (NPV) of 92% (Fig. 1 and Table 1). The area under Receiver Operating Characteristic (ROC) was 88% (Fig. 1C). The second classifier discriminated between noniPD patients and healthy subjects with 94% sensitivity, 81% specificity, and 97% NPV in the training set samples of known identity (Fig. 1D and Table 1). In the validation set this comparison scored 81% sensitivity, 76% specificity, and 90% NPV. The area under ROC was 85% (Fig. 1E). The confounding effect of L-dopa or MAO-B inhibitor treatment was tested by applying the first classifier on all patients, trying to distinguish between de-novo and L-dopa or MAOB inhibitor-treated parkinsonian patients. The results of this test showed only a random classification of the samples with ROC AUC of 56% and 59%, respectively (Fig. 1F). It is noteworthy to point out that the signals obtained when analyzing parkinsonian samples were more variant and less centralized in comparison to iPD or healthy control samples. This variance was also expressed in the relatively low PPV (62%) in the classifier validation, and could be explained by the heterogeneity of this group, which contains several sub-types: Drug Induced Parkinsonism (25%), Progressive Supranuclear Palsy (19%), Multiple System Atrophy (13%), CorticoBasal Ganglionic Degeneration (6%), Diffuse Lewy Body Disease (13%) and combined sub-types (25%) (Supplementary Table S1). On the other hand, the iPD group was much more homogenous, enabling better detection ability of the classifier (NPV ¼ 92%). 4. Discussion In this study we report on a preliminary result indicating that discrimination between iPD and non-iPD parkinsonism could be achieved by breath analysis using an array of nanomaterials based gas sensors with an accuracy of 81%. However, a potentially obvious limitation of the current study is that the diagnosis was based upon the specialist examination only, which could lack accuracy, instead
M.K. Nakhleh et al. / Parkinsonism and Related Disorders 21 (2015) 150e153
of the gold standard brain autopsy or by a combination of imaging techniques. On the other hand, periodic re-examination of the cohort has been carried out for 3 years, in order to monitor disease progression, development of clinical symptoms and the response to treatment, which together are used as confirmation of the initial diagnosis. The preliminary results presented here indicate a powerful ability of breath analysis to distinguish between non-iPD, iPD and healthy subjects. These results are comparable to transcranial sonography, which can discriminate between iPD and PSP or MSA, with 91% sensitivity and 82% specificity. However, the nanoarray has a great advantage of being more accurate in distinguishing between iPD and other parkinsonian syndromes [11]. The nanoarray has lower sensitivity (88%) than that of the olfactory disturbance among iPD cases (96%). Nevertheless, the nanoarray has higher specificity (81%) than that of the olfactory disturbance (63%) when the discrimination between iPD and MSA is targeted [12]. The alteration in the response of the nanoarray to exhaled air of parkinsonism compared to normal subjects is due to the overall changes in the content of exhaled breath volatile biomarkers. This observation was supported via parallel analysis with gaschromatography/mass spectrometry (GCeMS), which revealed tentatively 4 volatile organic compounds for the discrimination between iPD and non-iPD states (p-value < 0.05): 3methylhexane (11 ppb vs. 10 ppb), 2-pentanone (19 ppb vs. 4 ppb), benzaldehyde (31 ppb vs. 27 ppb); acetophenone (13 vs. 11 ppb), and 2,6 dimethylnonane. The same GCeMS analysis revealed three volatile biomarkers for the discrimination between parkinsonism and healthy states (p-value < 0.05): acetophenone (11 ppb vs. 13 pbb); 3-methyl hexane (10 ppb vs. 11 ppb); and 2,6-dimethylnonane. These significant alterations could be due to the effect of one or more of the following processes: oxidative stress, cytochrome p450, hepatic enzymes, carbohydrate-metabolism (glycolysis or gluconeogenesis pathways), lipid metabolism and others associated with the different neurological pathologies [6]. The origin and pathway of the reported compounds are now under extensive studies. Previously, it was evidenced that the nanoarray system is not influenced by potential confounding factors, such as age, gender, smoking and residence location [13,14]. In the current study, the results show no confounding influence of the L-dopa and/or MAO-B treatment on the performance of the applied nanoarray. This is because the nanoarray contains cross-reactive sensors that respond to most/all compounds in a mixture [9], rather than towards a specific biomarker. Therefore, any change in a single volatile marker of a given confounding factor would have negligible effect on the sensing signal, compared to the signal generated by the overall volatile compounds in the breath sample. In summary, we show that exhaled breath analysis with nanoarray is a promising non-invasive technique for distinguishing between different Parkinsonian states. Further development of the technique can produce a portable and inexpensive device which could be available at the level of individual clinics, however a larger clinical study is needed to validate the current results. Study funding This study was supported by the Kamin Fund, Israel Ministry of Trade and Industry (grant no. 47247).
153
Authors contribution Dr. Morad Nakhleh e Statistical analysis, interpretation of data and manuscript drafting. Final approval of the version to be submitted. Dr. Samih Badarny e Study design and supervision, and manuscript revision. Final approval of the version to be submitted. Dr. Raz Winer e Acquisition of clinical samples and data, revision of the manuscript for important intellectual content. Mrs. Raneen Jeries e Acquisition of clinical samples and data interpretation. Prof. John Finberg e Data interpretation and manuscript revision. Final approval of the version to be submitted. Prof. Hossam Haick e Study design and supervision, interpretation of data and manuscript revision. Final approval of the version to be submitted.
Financial disclosure All co-authors report no financial disclosures.
Appendix A. Supplementary data Supplementary data related to this article can be found at http:// dx.doi.org/10.1016/j.parkreldis.2014.11.023.
References [1] Hughes AJ, Daniel SE, Ben-Shlomo Y, Lees AJ. The accuracy of diagnosis of parkinsonian syndromes in a specialist movement disorder service. Brain 2002;125:861e70. [2] Schapira H. Recent developments in biomarkers in Parkinson disease. Curr Opin Neurol 2013;26:395e400. [3] Schrag A, Ben-Shlomo Y, Quinn N. How valid is the clinical diagnosis of Parkinson's disease in the community? J Neurol Neurosurg Psychiatry 2002;73: 529e34. [4] Berti V, Pupi A, Mosconi L. PET/CT in diagnosis of movement disorders. Ann N Y Acad Sci 2011. [5] Vogt T, Kramer K, Gartenschlaeger M, Schreckenberger M. Estimation of further disease progression of Parkinson's disease by dopamin transporter scan vs clinical rating. Parkinsonism Relat Disord 2011;17:459e63. [6] Haick H, Broza YY, Mochalski P, Ruzsanyi V, Amann A. Assessment, origin, and implementation of breath volatile cancer markers. Chem Soc Rev 2014;43: 1423e49. [7] Tisch U, Aluf Y, Ionescu R, Nakhleh M, Bassal R, Axelrod N, et al. Detection of asymptomatic nigrostriatal dopaminergic lesion in rats by exhaled air analysis using carbon nanotube sensors. ACS Chem Neurosci 2011;3:161e6. [8] Tisch U, Schlesinger I, Ionescu R, Nassar M, Axelrod N, Robertman D, et al. Detection of Alzheimer's and Parkinson's disease from exhaled breath using nanomaterial-based sensors. Nanomedicine (Future Medicine) 2013;8:43e56. [9] Broza YY, Haick H. Nanomaterial-based sensors for detection of disease by volatile organic compounds. Nanomedicine (Future Medicine) 2013;8: 785e806. [10] Breteton R. Chemometrics, application of mathematics and statistics to laboratory systems. J Chemometr 1990;6:228. [11] Berg D, Godau J, Walter U. Transcranial sonography in movement disorders. Lancet Neurol 2008;7:1044e55. [12] Haehner A, Hummel T, Reichmann H. Olfactory loss in Parkinson's disease. Parkinsons Dis 2011;450939:21. [13] Cohen-Kaminsky S, Nakhleh M, Perros F, Montani D, Girerd B, Garcia G, et al. A proof of concept for the detection and classification of pulmonary arterial hypertension through breath analysis with a sensor array. Am J Respir Crit Care Med 2013;188:756e9. [14] Nakhleh MK, Jeries R, Gharra A, Binder A, Broza YY, Pascoe M, et al. Detecting active pulmonary tuberculosis with a breath test using nanomaterial-based sensors. Eur Respir J 2014;43:1522e5.