INT J TUBERC LUNG DIS 17(5):682–686 © 2013 The Union http://dx.doi.org/10.5588/ijtld.12.0829
Artificial neural network models to support the diagnosis of pleural tuberculosis in adult patients J. M. Seixas,* J. Faria,* J. B. O. Souza Filho,† A. F. M. Vieira,‡ A. Kritski,§ A. Trajman‡¶ * Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia/Poli, Federal University of Rio de Janeiro, Rio de Janeiro, Rio de Janeiro, † Federal Centre of Technological Education Celso Suckov da Fonseca, Rio de Janeiro, Rio de Janeiro, ‡ Health Education Post-graduation Program, Gama Filho University, Rio de Janeiro, Rio de Janeiro, § Tuberculosis Academic Program, Medical School, Federal University of Rio de Janeiro, Rio de Janeiro, Rio de Janeiro, Brazil; ¶ McGill University Medical School, Montreal, Quebec, Canada SUMMARY BACKGROUND:
Clinicians in countries with high tuberculosis (TB) prevalence often treat pleural TB based on clinical grounds, as the availability and sensitivity of diagnostic tests are poor. O B J E C T I V E : To evaluate the role of artificial neural networks (ANN) as an aid for the non-invasive diagnosis of pleural TB. These tools can be used in simple computer devices (tablets) without remote internet connection. M E T H O D S : The clinical history and human immunodeficiency virus (HIV) status of 137 patients were prospectively entered in a database. Both non-linear ANN and the linear Fisher discriminant were used to calculate performance indexes based on clinical grounds. The same procedure was performed including pleural fluid test results (smear, culture, adenosine deaminase, serology
and nucleic acid amplification test). The gold standard was any positive test for TB. R E S U LT S : In pre-test modelling, the neural model reached >90% accuracy (Fisher discriminant 74.5%). Under pre-test conditions, ANN had better accuracy compared to each test considered separately. C O N C L U S I O N S : ANN are highly reliable for diagnosing pleural TB based on clinical grounds and HIV status only, and are useful even in remote conditions lacking access to sophisticated medical or computer infrastructure. In other better-equipped scenarios, these tools should be evaluated as substitutes for thoracocentesis and pleural biopsy. K E Y W O R D S : pleurisy; accuracy; artificial intelligence; tuberculosis; diagnosis
PLEURAL TUBERCULOSIS (pTB) is the second most frequent clinical presentation of TB, and a common cause of pleural effusion in many countries.1 The diagnosis of pTB is a challenge, as patients usually present with non-productive cough. Sputum is thus not available for acid-fast bacilli (AFB) staining and Mycobacterium tuberculosis culture, and pleural specimens are frequently negative for both tests.2 Pleural biopsy is frequently required for a rapid diagnosis of TB, usually by histopathological examination, which has 80% sensitivity.2 Problems with histopathological examination include lack of specificity and the increased risk of complication and costs added by pleural biopsy.3 Newer tests have been proposed to increase the probability of diagnosing pTB using pleural fluid specimens in adult subjects. These include non-specific inflammatory and immune response biomarkers (adenosine deaminase [ADA], neopterin, leptin, lysozyme, cytokines, complement activation, cell subsets) as well as specific markers of the immune response to
M. tuberculosis (T-cell response to specific antigens, B-cell response/antibody detection) and finally, detection of M. tuberculosis using nucleic acid amplification testing (NAAT).4 Scoring systems based on a combination of clinical variables and biomarkers have also been proposed as a diagnostic aid for this condition.4–12 Among the non-specific inflammatory biomarkers that have been evaluated, ADA and interferon-gamma (IFN-γ) are accepted as diagnostic aids and are the most accurate for pTB diagnosis, with consistently high sensitivity in many studies, although both are only expressions of the inflammatory process and not of TB aetiology.13 In a previous study, we observed that the measurement of ADA activity was the only test that had significantly higher sensitivity than histopathological examination.14 It is simple and inexpensive to perform and does not require special equipment, making it an attractive option for limitedresource settings with high TB incidence where the pre-test probability is high. Serological tests using
Correspondence to: A Trajman, Tuberculosis Academic Program, Federal University of Rio de Janeiro, Medical School, Rua Macedo Sobrinho 74/203, Rio de Janeiro 22271-080, Brazil. Tel: (+55) 21 2539 9194. Fax: (+55) 21 2532 1661. e-mail:
[email protected] Article submitted 29 October 2012. Final version accepted 18 December 2012.
Neural network for pleural TB diagnosis
antibodies against mycobacterial protein or glycolipids and commercial NAATs for M. tuberculosis detection have a potential role in confirming pleural TB due to their high specificity; however, these tests have low sensitivity and were therefore not useful in excluding the disease.14 Innovative approaches that take into account only clinical features and simple, inexpensive, rapid diagnostic tests should be developed with the use of modelling tools. Logistic regression models for scoring systems based on simple clinical and pleural fluid data have also shown high accuracy for pTB diagnosis in a few studies.5–7 These models are highly relevant to clinicians practising in resource-poor settings in high TB burden countries. Even when ADA is unavailable, the use of demographic and clinical information with or without a differential cell count might yield a pTB diagnosis with a high predictive value, particularly in high-prevalence settings.7 When handling high-dimensional classification problems, different modelling approaches may be used. Earlier works have applied multivariate logistic regression,15–17 classification trees18,19 and artificial neural networks (ANN)20,21 for predicting smearnegative pTB. The main advantage usually provided by ANN is their capability to extract hidden linear and non-linear relationships, even in highly dimensional and complex data sets.22 The objective of the present paper was to evaluate the use of ANN systems as an aid for a non-invasive diagnosis of pTB, based both on clinical grounds and on pleural fluid tests, to avoid or postpone the need for pleural biopsy. To better evaluate the gain in performance obtained from a non-linear model, the well-known Fisher linear model23 is also applied to these data.
METHODS Database The results from this work were obtained from the same data set used in previous studies by Trajman et al.,14 which assessed the diagnostic value of pleural fluid tests (ADA activity measurement, immunoglobulin A-enzyme-linked immunosorbent assay [IgAELISA] and NAAT) in the diagnosis of pTB. The study was approved by the Brazilian National Ethical Committee (#572/09). All patients with pleural effusion admitted for diagnostic purposes at the Santa Casa da Misericórdia Hospital, Rio de Janeiro, were eligible for the study; all participants provided signed, informed consent. Each patient record consisted of 10 symptoms or test result variables (plus age). These 10 variables were coded as +1 (affirmative), −1 (negative) and 0 (in case the specified test was not available or when the patient failed to provide information). The variables were derived from three groups: anamnesis
683
(age, sex [+1 for male], smoker); classical tests (histopathological, pleural tissue culture, pleural fluid culture, AFB smear, HIV testing), and ‘newer’ tests such as ADA, IgA-ELISA and NAAT. The database contains records for 137 patients, two of whom were excluded because their records lacked a final diagnosis. Of the 135 remaining patients, 96 were diagnosed with pTB, of whom 22 were diagnosed solely on clinical grounds: none of the classical tests was positive.14 Design The performance of the classifiers was measured according to sensitivity, specificity and accuracy. In addition, the sum-product (SP) index (ISP)24 was considered. This index is computed as the geometrical mean of both geometrical and arithmetical means of sensitivity and specificity: ISP =
√ S +2 E √ S × E
(1)
where S is sensitivity and E is specificity, both in percentage. The SP index is suitable for balanced classifier designs with respect to both sensitivity and specificity, as the geometrical mean forces the index to drop drastically when either sensitivity or specificity is significantly reduced. Simulations were performed based on two approaches: pre-test and post-test. For pre-test evaluation, only anamnesis variables and HIV status were used. This produces a score that can be accessed immediately after patient interviews and a point-of-care test (HIV), which may be important for immediate treatment initiation in any setting, including those with poor laboratory infrastructure. For post-test evaluation, in addition to the anamnesis variables, ‘new’ and classical test variables were used, excluding histopathological and pleural tissue culture, as these are only available upon biopsy. As the proposal of the diagnosis support system is to provide a fast, non-invasive score, results from these two last tests were ignored. For post-test evaluations, ANN provide data fusion models incorporating diverse information from patients. Neural network design The neural pTB models used a two-layer feedforward multilayer perceptron (MLP) topology (Figure 1), which has universal approximation property, i.e., it can approximate any arbitrary continuous mapping function.22 To produce highly accurate prediction models, several network design parameters were carefully investigated, with emphasis on the choice of the number of hidden layer neurons, as well as aspects related to the training process. These parameters were defined using a cross-validation approach25 based on the SP index.
684
The International Journal of Tuberculosis and Lung Disease
Figure 1 A sketch of the multilayer perceptron (MLP) model, showing how input information is processed and how the artificial neuron is implemented.
The cross-validation was based on a stratified random subsampling method.26 The data set was thus randomly split into different training and test sets. The former was used for determining network parameters (weights and biases) and the latter for estimating network performance (training procedure generalisation). For each data split, MLP networks with from 1 to 20 hidden neurons were produced. Each network was trained 100 times using different initial random values to avoid local minimum problem.22 The best set of initialisation parameters was selected according to the SP index. To ensure good generalisation performance (avoiding overlearning), the early stop method was adopted as the learning mechanism control.22 This process consists of interrupting the training phase when a loss in generalisation is observed due to overtraining. Early stop procedure was based on computing the mean-squared error (MSE) between target and output values from test set samples. Performance was estimated using the average and standard deviation of the SP index attained over all runs (100 in this study). These splits considered test sets containing 10% of the data set size. The number
of hidden neurons was determined based on the maximum performance value obtained from the test set. Age values were normalised to attain zero mean and extreme values of +1 and −1 in the training group. Age values in the test groups were then normalised using the parameters obtained from the corresponding training groups. The training algorithm adopted was resilient backpropagation (RPROP), operated in batch mode.27 RPROP provides nearly optimal convergence time with little extra computational expense and need for parameter tweaking. The MSE was the cost function, and all neurons employed the hyperbolic tangent as their activation function. The target values for training were +1 for a positive pTB diagnosis sample and −1 otherwise. For the purposes of comparison, the performance of alternative diagnostic methods, including combinations of these, were similarly evaluated. The combination criterion was to consider a positive diagnosis if any method showed a positive outcome. As with ANN models, performance was also evaluated using the cross-validation approach.
RESULTS Four models were produced, two for each diagnosis support approach (pre-test and post-test). The results obtained (and their 95% confidence interval [CI]) for the models, all selected by the highest average SP index value, are outlined in Table 1. During pre-test, the non-linear model (neural network based) achieved an average SP index value of >90%. As shown in Figure 2, the neural model average performance is above the performance achieved by any non-preprocessed data. Moreover, ANN identified all 22 patients diagnosed solely on clinical grounds. The results of post-test modelling are shown in Table 2, and are compared to both Fisher and neural models in Figure 2. The non-linear model SP index performance exceeded that of any single or combination test result.
DISCUSSION Early diagnosis and treatment of pTB, a common form of presentation of the disease, is important as it reduces pleural sequelae and progression to pulmonary
Table 1 ‘Selected models’ performance estimates for the diagnosis of tuberculosis in a data set of 135 patients with pleural effusion; performance outcomes of the test set
Technique
Approach
Hidden neurons n
Fisher’s Fisher’s ANN ANN
Pre-test Post-test Pre-test Post-test
— — 15 18
Sensitivity % (95%CI)
Specificity % (95%CI)
Accuracy % (95%CI)
SP % (95%CI)
68.2 (66.6–69.8) 93.0 (92.4–93.6) 94.5 (91.4–97.6) 99.3 (98.0–100.0)
90.2 (88.6–91.8) 94.9 (94.2–95.6) 91.0 (85.2–96.8) 99.2 (97.3–100.0)
74.5 (73.6–75.4) 93.5 (93.1–93.9) 93.5 (91.0–96.0) 99.3 (98.2–100.0)
78.8 (78.2–79.4) 93.9 (93.5–94.3) 92.6 (89.5–95.7) 99.3 (98.1–100.0)
CI = confidence interval; SP = sum-product index; ANN = artificial neural network.
Neural network for pleural TB diagnosis
685
Figure 2 Results in terms of sensitivity (A) and specificity (B) from neural networks and Fisher linear discriminator (for both preand post-test), and different tests or combination of tests. Open squares represent mean values and their respective error bars are also shown. ANN = artificial neural network; FLD = Fisher linear discriminant; AFB = acid-fast bacilli; ADA = adenosine deaminase; ELISA = enzyme-linked immunosorbent assay; PCR = polymerase chain reaction.
disease. The diagnosis of pTB is frequently based on clinical grounds or histopathological evidence suggestive of, but not pathognomonic for, TB. Non-specific biomarkers of the inflammatory process have also been used as surrogates of TB. Nevertheless, pleural tissue specimens require an infrastructure for performing biopsies, and biochemical, immunological and NAAT tests are rarely available in remote settings. We thus propose the development of a pTB diagnostic support system. In this article we have shown that the performance attained by the ANN model for the diagnosis of pTB was equivalent to the histopathological test and to reliable ADA findings. Although these are not considered definitive evidence of tuberculous origin, they are the criteria most frequently used for starting treatment. This support system raised the 70% pre-test probability (prevalence) to over 90%. In other words, the
clinician could have a high degree of confidence to treat TB without performing any other tests. Adding simple biochemical tests raised the probability further to 99%, retaining high specificity. The proposed predictor system is therefore useful for ruling in and ruling out TB, and it could provide continuous support during the different steps prior to the final diagnosis, with different models analysing patient data in the first moment of suspicion of pTB and, at a later point, when pleural fluid test results are available. It could therefore be used for two different functions: by helping with the logistics of patient management in remote places with limited access to laboratory resources, and by avoiding the use of pleural biopsy, thus reducing the morbidity and health care costs associated with these procedures. However, caution is needed when generalising these findings to low TB prevalence settings. As for any other test, according to
Table 2 Variable performance estimates for the diagnosis of tuberculosis in a data set of 135 patients with pleural effusion; performance outcomes of the test set used in ANN evaluation Variable Histopathological Tissue culture Fluid culture AFB smear ADA ELISA NAAT ADA+ELISA ELISA+NAAT ADA+NAAT ADA+ELISA+NAAT
Sensitivity % (95%CI)
Specificity % (95%CI)
Accuracy % (95%CI)
SP % (95%CI)
73.2 (66.7–79.7) 14.3 (9.9–18.7) 6.6 (3.3–9.9) 0.5 (0.0–1.5) 79.9 (74.8–85.0) 59.6 (52.1–67.1) 66.9 (60.4–73.4) 88.6 (84.4–92.8) 80.4 (75.7–85.1) 85.4 (81.0–89.8) 92.6 (89.1–96.1)
100.0 (100.0–100.0) 100.0 (100.0–100.0) 100.0 (100.0–100.0) 100.0 (100.0–100.0) 94.0 (89.2–98.8) 98.2 (97.3–100.0) 90.5 (84.0–97.0) 91.8 (86.5–97.1) 87.2 (79.8–94.6) 83.8 (75.5–92.1) 79.0 (70.3–87.7)
80.9 (76.3–85.5) 38.8 (35.6–42.0) 33.3 (31.0–35.6) 28.9 (28.2–29.6) 83.9 (80.2–87.6) 70.6 (65.3–75.9) 73.6 (68.6–78.6) 89.5 (86.2–92.8) 82.4 (78.0–86.8) 84.9 (80.9–88.9) 88.7 (85.3–92.1)
85.9 (82.3–89.5) 40.0 (31.1–48.9) 23.1 (13.0–33.2) 0.2 (0.0–4.3) 86.6 (83.2–90.0) 77.2 (72.7–81.7) 77.8 (72.9–82.7) 90.0 (86.7–93.3) 83.6 (78.7–88.5) 84.2 (79.3–89.1) 85.2 (80.3–90.1)
ANN = artificial neural network; CI = confidence interval; SP = sum-product index; AFB = acid-fast bacilli; ADA = adenosine deaminase; ELISA = enzymelinked immunosorbent assay serology for IgA to specific M. tuberculosis antigens;14 NAAT = nucleic acid amplification technique.
686
The International Journal of Tuberculosis and Lung Disease
the Bayesian approach, the positive predictive value is higher in high TB prevalence settings, and this system should be further investigated in other scenarios. A major limitation of the present study is the low number of control patients in the database, which affects the assessment of specificity and, in turn, the ability to rule out other diseases, the main concern being to rule out cancer, in particular among smokers. However, these results provide a first evaluation of the potentials and challenges faced when dealing with the support of pTB diagnosis. In addition, there is no clear gold standard for the diagnosis of pTB, as culture and NAAT tests, which are evidence of the presence of an aetiological agent, have very low sensitivities. Future studies may include the refinement of these methods using a larger database, now under construction, and the development of new models based on different techniques and approaches, especially models based on clustering techniques, such as the adaptive resonance theory,22 for group risk assignment of patients. These models, used as a simple application in tools that do not require an internet connection, such as mobile phones or tablets, could simplify the diagnostic process and give support for the rapid initiation of treatment in distant settings, with a consequent reduction in unnecessary costs and delays.
11
12
13
14
15
16
17
18
Acknowledgements The authors thank Fundação de Amparo À Pesquisa do Estado do Rio de Janeiro and Conselho Nacional de Desenvolvimento Científico e Tecnológico (Brazil) for their support. Conflict of interest: none declared.
References 1 Valdés L, Pose A, San José E, Martinez Vázquez J M. Tuberculous pleural effusions. Eur J Intern Med 2003; 14: 77–88. 2 Light R. Establishing the diagnosis of tuberculous pleuritis. Arch Intern Med 1998; 158: 1967–1968. 3 Berger H W, Mejia E. Tuberculous pleurisy. Chest 1973; 63: 88–92. 4 Trajman A, Pai M, Dheda K, Van Zul Smit R, et al. Novel tests for diagnosing tuberculosis effusion: what works and what does not? Eur Respir J 2008; 31: 1098–1106. 5 Villegas M V, Labrada L A, Saravia N G. Evaluation of polymerase chain reaction, adenosine deaminase and interferongamma in pleural fluid for the differential diagnosis of pleural tuberculosis. Chest 2000; 118: 1355–1364. 6 Porcel J M, Vives M. Differentiating tuberculous from malignant pleural effusions: a scoring model. Med Sci Monit 2003; 9: CR175–180. 7 Neves D D, Dias R M, Cunha A J. Predictive model for the diagnosis of tuberculous pleural effusion. Braz J Infect Dis 2007; 11: 83–88. 8 Greco S, Girardi E, Masciangelo R, Capoccetta G B, Saltini C. Adenosine deaminase and interferon gamma measurements for the diagnosis of tuberculous pleurisy: a meta-analysis. Int J Tuberc Lung Dis 2003; 7: 777–786. 9 Pai M, Flores L L, Hubbard A, Riley L W, Colford J M Jr. Nucleic acid amplification tests in the diagnosis of tuberculous pleuritis: a systematic review and meta-analysis. BMC Infect Dis 2004; 4: 6–19. 10 Jiang J, Shi H Z, Liang Q L, Qin S M, Qin X J. Diagnostic
19
20
21
22 23 24
25 26
27
value of interferon-c in tuberculous pleurisy: a meta-analysis. Chest 2007; 131: 1133–1141. Sales R K, Vargas F S, Capelozzi V L, et al. Predictive models for diagnosis of pleural effusions secondary to tuberculosis or cancer. Respirology 2009; 14: 1128–1133. Valdés L, San José M E, Pose A, et al. Diagnosing tuberculous pleural effusion using clinical data and pleural fluid analysis. A study of patients less than 40 years-old in an area with a high incidence of tuberculosis. Respir Med 2010; 104: 1211–1217. Kalantri Y, Hemvani N, Chitnis D S. Evaluation of real-time polymerase chain reaction, interferon-gamma, adenosine deaminase, and immunoglobulin A for the efficient diagnosis of pleural tuberculosis. Int J Infect Dis 2011; 15: 226–231. Trajman A, Kaisermann C, Luiz R R, et al. Pleural fluid ADA, IgA-ELISA and NAAT sensitivities for the diagnosis of pleural tuberculosis. Scand J Clin Lab Invest 2007; 67: 877– 884. Ruel M, Bouvet E, Tattevin P, Casalino E, Fleury L, Egmann G. The validity of medical history, classic symptoms, and chest radiographs in predicting pulmonary tuberculosis: derivation of a pulmonary tuberculosis prediction model. Chest 1999; 115: 1248–1253. Wisnivesky J P, Serebrisky D, Moore D, Sacks H S, Iannuzzi M C, McGinn T. Validity of clinical prediction rules for isolating inpatients with suspected tuberculosis. A systematic review. J Gen Intern Med 2005; 20: 947–952. Solari L, Acuna-Villaorduna C, Soto A, Agapito J, et al. A clinical prediction rule for pulmonary tuberculosis in emergency departments. Int J Tuberc Lung Dis 2008; 12: 619–624. Mello F C, Bastos L G, Soares S L, et al. Predicting smearnegative pulmonary tuberculosis with classification trees and logistic regression: a cross-sectional study. BMC Public Health 2006; 6: 43–50. Cain K P, McCarthy K D, Heilig C M, et al. An algorithm for tuberculosis screening and diagnosis in people with HIV. N Engl J Med 2010; 362: 707–716. El-Solh A A, Hsiao C B, Goodnough S, Serghani J, Grant B J. Predicting active pulmonary tuberculosis using an artificial neural network. Chest 1999; 116: 968–973. Seixas J M, Mello F C Q, Santos A M, Pereira B B, Kristski A L. Neural networks: an application for predicting smear negative pulmonary tuberculosis. In: Balakrishnan N, Auget J L, Mesbah M, Molenberghs G. Advances in statistical methods for the health sciences. Boston, MA, USA: Birkhäuser, 2006: pp 279–292. Haykin S. Neural networks and learning machines. 3rd ed. Upper Saddle River, NJ, USA: Pearson Education Inc, 2009. Duda R, Hart P, Stork D. Pattern classification. New York, NY, USA: Wiley, 2000. de Simas Filho E F, Seixas J M, Calôba L P. Online neural filtering operating over segmented discriminating components. Proceedings of the 15th IEEE International Conference on Electronics, Circuits and Systems, 31 August–3 September 2008, St. Julien’s, Malta. Washington DC, USA: IEEE, 2008: pp 530– 533. http://ieeexplore.ieee.org/xpl/tocresult.jsp?sortType%3 Dasc_p_Sequence%26filter%3DAND%28p_IS_Number% 3A4674773%29&searchWithin=simas&pageNumber=1& resultAction=REFINE Accessed February 2013. Efrom B, Tibshirani R J. An introduction to the bootstrap. New York, NY, USA: Chappman & Hall/CRC, 1994. Kohavi R. A study of cross-validation and bootstrap for accuracy estimation and model selection. Proceedings of International Joint Conference on Artificial Intelligence, 20–25 August 1995, Montreal, QC, Canada. San Francisco, CA, USA: Morgan Kaufmann, 1995: 1137–1143. Riedmiller M. Rprop—description and implementation details. Technical report. Karlsruhe, Sweden: Institut fur Ligik, University of Karlsruhe, 1994.
Neural network for pleural TB diagnosis
i
RÉSUMÉ
: Dans les pays à haute incidence de tuberculose (TB), les cliniciens traitent fréquemment la TB pleurale en se basant sur des données cliniques, car la disponibilité et la sensibilité des tests de diagnostic sont médiocres. O B J E C T I F : Evaluer dans ce contexte le rôle des réseaux de neurones artificiels (ANN) comme aide au diagnostic non-invasif de la TB pleurale. Ces outils peuvent être utilisés dans certains appareils simples d’informatique (tablettes) sans connection avec un réseau éloigné. M É T H O D E S : On a introduit de manière prospective dans une base de données l’histoire clinique et le statut du virus de l’immunodéficience humaine (VIH) de 137 patients. On a utilisé pour calculer les index de performance basés sur les données cliniques à la fois l’ANN non-linéaire et le discriminant linéaire de Fisher. La même procédure a été réalisée en y incluant les résultats des tests portant sur le liquide pleural (frottis, culture, CADRE
adénosine déaminase, examens sérologiques et test d’amplification des acides nucléiques). On a considéré comme gold standard n’importe quel test positif pour la TB. R É S U LTAT S : Dans les modèles pré-test, le modèle neuronal a atteint plus de 90% de précision (contre 74,5% pour le discriminant de Fisher). Dans les conditions prétest, la précision de l’ANN est meilleure par comparaison avec chacun des tests séparés. C O N C L U S I O N S : L’ANN est très fiable pour le diagnostic de la TB pleurale reposant sur des données cliniques et sur le statut VIH ; il est très utile même dans des conditions éloignées sans accès à une infrastructure médicale ou informatique sophistiquée. Dans d’autres scénarios mieux équipés, ces outils devraient être évalués comme substitut à la thoracocentèse et à la biopsie pleurale.
RESUMEN M A R C O D E R E F E R E N C I A : Los médicos en los países con alta prevalencia de tuberculosis (TB) suelen tratar la TB pleural con base en las características clínicas, debido a la escasez de pruebas diagnósticas y a su baja sensibilidad. O B J E C T I V O : Evaluar la utilidad de las redes neurales artificiales (ANN) como una ayuda diagnóstica no invasivo de la TB pleural. Estos instrumentos se pueden utilizar en dispositivos informáticos sencillos (tabletas) sin conexión remota a una red. M É T O D O S : Se incorporaron a una base de datos de manera prospectiva, los datos provenientes de la anamnesis y el resultado del examen serológico frente al virus de la inmunodeficiencia humana (VIH) de 137 pacientes. Con el fin de calcular los índices de rendimiento se utilizaron los métodos no lineares de las ANN y el análisis discriminatorio lineal de Fisher, con base en los criterios clínicos. Se llevó a cabo el mismo procedimiento a partir de los resultados del examen del líquido pleural
(frotis, cultivo, actividad de la desaminasa de adenosina, la serología VIH y las pruebas de amplificación de ácidos nucleicos). El criterio de referencia fue toda prueba positiva de TB. R E S U LTA D O S : En la simulación con modelos antes de la prueba, el modelo neural alcanzó una exactitud superior a 90% (74,5% con el análisis discriminatorio de Fisher). En las condiciones previas a la investigación el método de ANN ofreció una mayor exactitud, al compararlo con cada prueba separadamente. C O N C L U S I Ó N : El método de las ANN es sumamente fiable en el diagnóstico de la TB pleural con base en los criterios clínicos y la serología VIH exclusivamente y se puede utilizar incluso en entornos remotos, donde se carece de acceso a infraestructuras médicas o informáticas sofisticadas. En medios con dotaciones más completas, se debería evaluar la utilidad de estos instrumentos como sustitutos de la toracocentesis y la biopsia pleural.
