Effect of incremental feature enrichment on healthcare text classification system: A machine learning paradigm

Healthcare tweets are particularly challenging due to its sparse layout and its limited character size. Compared to previous method based on "bag of words" (BOW) model, this study uniquely identifies the enrichment protocol and learns how semantically different aspects of feature selection such as BOW (feature F0), term frequency inverse document frequency (TF-IDF, feature F1), and latent semantic indexing (LSI, feature F2) when applied sequentially with classifier improves the overall performance.

To study this enrichment concept, our ML model is tested on two kinds of diverse data sets: (i) D1: Disease data with conjunctivitis, diarrhea, stomach ache, cough and nausea related tweets, and (ii) D2: WebKB4 dataset, while adapting three kind of classifiers (a) C1: support vector machine with radial basis function (SVMR), (b) C2: Multi-layer perceptron (MLP) and (c) C3: Random Forest (RF). Partition protocol (K10) was adapted with different performance metrics to evaluate machine learning (ML)-system.

Using the combination of F1, C1, D1, K10, ML accuracy was: 94%, while with F2, C1, D1, K10, ML accuracy was 97%. Using the incremental feature enrichment from F0 to F2, K10 protocol gave F1 improvement over F0 by 4.98% on Disease dataset, while F2 improvement over F0 was by 11.78% on WebKB4 dataset. We demonstrated the generalization over memorization process in our ML-design. The system was tested for stability and reliability.

We conclude that semantically different aspects of feature selection, when adapted sequentially, leads to improvement in ML-accuracy for healthcare data sets. We validated the system by taking non-healthcare data sets.