Artificial neural networks help to identify disease subsets and to predict lymphoma in primary Sjögren's syndrome.

OBJECTIVES

Primary Sjögren's syndrome (pSS) is a complex chronic systemic disorder, for which specific and effective therapeutic interventions are still lacking. In this era of precision medicine, there is a clear need for a better definition of disease phenotypes to foster the research of novel specific biomarkers and new therapeutic targets. The main objectives of this work are: 1) to compare Auto Contractive Map (AutoCM), a data mining tool based on an artificial neural network (ANN) versus conventional Principal Component Analysis (PCA) in discriminating different pSS subsets and 2) to specifically focus on variables predictive of MALT-NHL development, assessing the previsional gain of the predictive models developed.

METHODS

Out of a historic cohort of 850 patients, we selected 542 cases of pSS fulfilling the AECG criteria 2002. Thirty-seven variables were analysed including: patient demographics, glandular symptoms, systemic features, biological abnormalities and MALT-NHLs. AutoCM was used to compute the association of strength of each variable with all other variables in the dataset. PCA was applied to the same data set.

RESULTS

Both PCA and AutoCM confirmed the associations between autoantibody positivity and several pSS clinical manifestations, highlighting the importance of serological biomarkers in pSS phenotyping. However, AutoCM allowed us to clearly distinguish pSS patients presenting with predominant glandular manifestations and no or mild extra-glandular features from those with a more severe clinical presentation. Out of 542 patients, we had 27 cases of MALT-NHLs. The AutoCM highlighted that, besides other traditional lymphoproliferative risk factors (i.e. salivary gland enlargement, low C4, leukocytopenia, cryoglobulins, monoclonal gammopathy, disease duration), rheumatoid factor was strongly associated to MALT-NHLs development. By applying data mining analysis, we obtained a predictive model characterised by a sensitivity of 92.5% and a specificity of 98%. If we restricted the analysis to the seven most significant variables, the sensitivity of the model was 96.2% and its specificity 96%.

CONCLUSIONS

Our study has shed new light on the possibility of using novel tools to extract hidden, previously unknown and potentially useful information in complex diseases like pSS, facing the challenge of disease phenotyping as a prerequisite for discovering novel specific biomarkers and new therapeutic targets.