Publications

(REVERSE-CHRONOLOGICALLY)

1. Mining ICDDR,B Hospital Surveillance Data Using Locally Linear Embedding based SMOTE Algorithm and Multilayer Perceptron. Adnan Firoze, Rashedur M Rahman.  'Lecture Notes in Intelligent Information and Database Systems'. Springer International Publishingpp 398-407. January, 2015. DOI: http://dx.doi.org/10.1007/978-3-319-15702-3_39

        Abstract: In this research, we have used a number of classifier models on Hospital Surveillance data to classify over 28,000 admitted patients according to their critical conditions. Three class labels were used to distinguish the criticality of the admitted patients. Furthermore, we have set forth two distinct approaches to address the over-fitting problem in the unbalanced dataset since the frequency of instances of the class ‘low’ is significantly higher than the other two classes. Apart from trimming the dataset to balance the classes, we have dealt with the over-fitting problem by introducing the ‘Synthetic Minority Over-sampling Technique’ (SMOTE) algorithm coupled with Locally Linear Embedding (LLE). We have constructed three models where we applied the neural, and multinomial logistic regression classifications and finally compared the performance of our models with the models developed by Rahman and Hasan (2011) where they used several decision tree models to classify the same dataset using tenfold cross validation. Additionally, for a comprehensive comparative analysis, we have compared the classification performance of our novel third model using support vector machine (SVM). After comparison, we have shown that one of our models surpasses all prior models in terms of classification performance, taking into account the performance time trade-off, giving us an efficient model that handles large scale unbalanced datasets efficiently with standard classification performance. The models developed in this research can become imperative tools to doctors when large numbers of patients arrive in a short interval especially during epidemics. Since, intervention of machines become a necessity when doctors are scarce, computer applications powered by these models are helpful to diagnose and measure the criticality of the newly arrived patients with the help of the historical data kept in the surveillance database.

2. Bangla User Adaptive Word Speech Recognition – Approaches and Comparisons. Adnan Firoze, M. Shamsul Arifin, Rashedur M. Rahman. International Journal of Fuzzy System Applications (IJFSA). Volume 3 Issue 3. 2013. DOI: http://dx.doi.org/doi:10.4018/ijfsa.2013070101 

        Abstract: The paper presents Bangla word speech recognition using two novel approaches developed by the authors and exhibits a comprehensive comparative analysis. The first approach is using spectral analysis and fuzzy logic and the other approach is based on Mel-frequency cepstral coefficients (MFCC) analysis and feed-forward back-propagation neural networks. As human speech is imprecise and ambiguous, fuzzy logic – the base of which is indeed linguistic ambiguity, could serve as a precise tool for analyzing and recognizing human speech. Even though the core source of an uttered word is a voiced signal, our systems revolve around the visual representations of voiced signals – the Fourier energy spectrum and the MFCC. The spectrogram may be perceived as a “visual” entity and on the other hand MFCC analysis is viewed as a warping of spectral analysis which emulates the responses of the human ear more closely. The essences of a Fourier energy spectrum and the MFCC are matrices that include information about properties of a sound by storing energy and frequency in discrete time. In this research, the spectral and cepstral analysis has been chosen as opposed to image processing for increased accuracy. The decision making process of our systems are based on fuzzy logic and neural networks. Experimental results (tenfold cross-validation over a sample size of one thousand utterances/word samples) demonstrate that our fuzzy logic based system is 86% accurate and our ANN based system is 90% accurate compared to a commercial Hidden Markov Model (HMM) based speech recognizer that shows 73% accuracy on an average. However, our research also derives that, even though ANN gives a better recognition accuracy than the fuzzy logic based system, the fuzzy logic based system is more accurate when it comes to “more difficult” or “polysyllabic” words (for which fuzzy logic based system shows a ten-fold cross validated accuracy of 90% whereas the ANN based system shows accuracy of 79%). In terms of runtime performance, the fuzzy logic based system surpasses the ANN based Bangla speech recognition system by demonstrating much less training time in our research.

3. Dermatological Disease Diagnosis using Color-skin Images. Shamsul Arifin, M.; Golam Kibria, M.; Firoze, Adnan; Ashraful Amin, M.; Hong Yan; Machine Learning and Cybernetics (ICMLC), 2012 International Conference on , vol.5, no., pp.1675-1680, 15-17 July 2012 doi: 10.1109/ICMLC.2012.6359626. (Full Text / PDF) (Summary in CVC Website)

Abstract: This paper presents an automated dermatological diagnostic system. Etymologically, dermatology is the medical discipline of analysis and treatment of skin anomalies. The system presented is a machine intervention in contrast to human arbitration into the conventional medical personnel based ideology of dermatological diagnosis. The system works on two dependent steps - the first detects skin anomalies and the latter identifies the diseases. The system operates on visual input i.e. high resolution color images and patient history. In terms of machine intervention, the system uses color image processing techniques, k-means clustering and color gradient techniques to identify the diseased skin. For disease classification, the system resorts to feedforward backpropagation artificial neural networks. The system exhibits a diseased skin detection accuracy of 95.99% and disease identification accuracy of 94.016% while tested for a total of 2055 diseased areas in 704 skin images for 6 diseases.

 4. Mining ICDDR, B Hospital Surveillance Data and Exhibiting Strategies for Balancing Large Unbalanced Datasets. Adnan Firoze, Rashedur M. Rahman.  International Journal of Healthcare Information Systems and Informatics (IJHISI)  10(1) 2015: 39-66. 

Abstract: This research uses a number of classifier models on Hospital Surveillance data to classify admitted patients according to their critical conditions. Three class labels were used to distinguish the criticality of the admitted patients. Furthermore, set forth are two distinct approaches to address the over-fitting problem in the unbalanced dataset since the frequency of instances of the class ‘low' is significantly higher than the other two classes. Apart from trimming the dataset to balance the classes, this work has dealt with the over-fitting problem by introducing the ‘Synthetic Minority Over-sampling Technique' (SMOTE) algorithm coupled with Locally Linear Embedding (LLE). It has constructed three models that applied the neural, and multinomial logistic regression classifications and finally compared the performance of the work's models with the models developed by Rahman and Hasan (2011) where they used several decision tree models to classify the same dataset using tenfold cross validation. Additionally, for a comprehensive comparative analysis, this work has compared the classification performance of the authors' novel third model using support vector machine (SVM). After comparison, the work shows that one of the authors' models surpasses all prior models in terms of classification performance, taking into account the performance time trade-off, giving them an efficient model that handles large scale unbalanced datasets efficiently with standard classification performance. The models developed in this research can become imperative tools to doctors when large numbers of patients arrive in a short interval especially during epidemics. Since, intervention of machines become a necessity when doctors are scarce, computer applications powered by these models are helpful to diagnose and measure the criticality of the newly arrived patients with the help of the historical data kept in the surveillance database.

 5. BANGLA Isolated Word Speech Recognition. Adnan Firoze, M. Shamsul Arifin, Ryana Quadir, Rashedur M. Rahman.  13th International Conference on Enterprise Information Systems (ICEIS) (Location: China)  (2) 2011: 73-82. (Full Text / PDF) (Citations)

Abstract: The paper presents Bangla word speech recognition using spectral analysis and fuzzy logic. As human speech is imprecise and ambiguous, the fuzzy logic – the base of which is indeed linguistic ambiguity, could serve as a more precise tool for analysing and recognizing human speech. Even though the core source of an uttered word is a voiced signal, our system revolves around the visual representation of voiced signals – the
spectrogram. The spectrogram may be perceived as a “visual” entity. The essences of a spectrogram are
matrices that include information about properties of a sound, e.g., energy, frequency and time. In this
research the spectral analysis has been chosen as opposed to image processing for increased accuracy. The
decision making process of our system is based on fuzzy logic. Experimental results demonstrate that our
system is 80% accurate compared to a commercial Hidden Markov Model (HMM) based speech recognizer
that shows 73% accuracy on an average.