Objective prediction of pharyngeal swallow dysfunction in dysphagia through artificial neural network modeling

BackgroundPharyngeal pressure-flow analysis (PFA) of high resolution impedance-manometry (HRIM) with calculation of the swallow risk index (SRI) can quantify swallow dysfunction predisposing to aspiration. We explored the potential use of artificial neural networks (ANN) to model the relationship between PFA swallow metrics and aspiration and to predict swallow dysfunction. MethodsTwo hundred consecutive dysphagia patients referred for videofluoroscopy and HRIM were assessed. Presence of aspiration was scored and PFA software derived 13 metrics and the SRI. An ANN was created and optimized over training cycles to achieve optimal classification accuracy for matching inputs (PFA metrics) to output (presence of aspiration on videofluoroscopy). Application of the ANN returned a value between 0.00 and 1.00 reflecting the degree of swallow dysfunction. Key ResultsTwenty one patients were excluded due to insufficient number of swallows (<4). Of 179, 58 aspirated and 27 had aspiration pneumonia history. The SRI was higher in aspirators (aspiration 24 [9, 41] vs no aspiration 7 [2, 18], p<0.001) and patients with pneumonia (pneumonia 27 [5, 42] vs no pneumonia 8 [3, 24], p<0.05). The ANN Predicted Risk was higher in aspirators (aspiration 0.57 [0.38, 0.82] vs no aspiration 0.13 [0.4, 0.25], p<0.001) and in patients with pneumonia (pneumonia 0.46 [0.18, 0.60] vs no pneumonia 0.18 [0.6, 0.49], p<0.01). Prognostic value of the ANN was superior to the SRI. Conclusions & InferencesIn a heterogeneous cohort of dysphagia patients, PFA with ANN modeling offers enhanced detection of clinically significant swallowing dysfunction, probably more accurately reflecting the complex interplay of swallow characteristics that causes aspiration.