In February of 2014, Dr. Paul Rodrigues and Dr. Karunanidhi Duraisamy together published a paper on machine learning-driven voice analysis for the diagnosis of various diseases. It was titled Analysis of Speech Algorithms in Disease affected Voice Patterns and appeared in the International Journal of Computer Science and Engineering, Volume 2, Issue 2.
In this paper's fourth section, it addresses Parkinson's disease. The section begins as follows:
"Parkinson’s disease is a neurodegenerative disorder of central nervous system that causes partial or full loss in motor reflexes, speech, behavior, mental processing, and other vital functions [1]. In 1817, PD was described as “shaking palsy” by Doctor James Parkinson [2]. It is generally observed in elderly people and causes disorders in speech and motor abilities (writing, balance, etc.) of 90% of the patients [3]."
Interestingly, this is, verbatim, down to the gramatical errors and citations, exactly how the 2013 paper Collection and Analysis of a Parkinson Speech Dataset With Multiple Types of Sound Recordings begins.
In fact, 100% of Rodrigues and Duraisamy's 2014 study's fourth section, "IV. ANALYSIS OF PARKINSON’S DISEASE" (including the inset figure, Fig. 1), is brazenly copied, unedited, from this 2013 paper.
Their paper's prior section, "III. COMPARISON OF SPEECH AND NON-SPEECH SOUNDS", is not. Instead, it begins by largely copying an initial list from the 2004 doctoral thesis of Dr. Michael Cowling, Non-Speech Environmental Sound Classification System for Autonomous Surveillance.
Plagiarised text
It was found that six techniques are commonly used for speech/speaker recognition or have been used for this domain in the past. These were:
- Dynamic Time Warping (DTW)
- Hidden Markov Models (HMM)
- Vector Quantization (VQ)
- Mel - Ceptral Co-efficient
- Artificial Neural Networks (ANN)
Original text
The following classification techniques are commonly used for speech/speaker recognition or have, in the past, been used for this application domain. They are:
- Dynamic Time Warping (DTW)
- Hidden Markov Models (HMM)
- Learning Vector Quantization (LVQ)
- Self-Organising Maps (SOM)
- Ergodic-HMM's
- Artificial Neural Networks (ANN)
- Long-Term Statistics (LTS)
It then copies phrasing from a small segment from the 2002 paper Analysis of Speech Recognition Techniques for use in a Non-Speech Sound Recognition System, of which Dr. Michael Cowling is the lead author.
Plagiarised text
Looking at these comparison tables, we can begin to examine whether any of these speech recognition or speaker identification techniques can be used for no speech sound recognition [5].
Original text
It then performs benchmarks on two of these techniques (LVQ and ANN’s) and determines which technique is better suited for non-speech sound recognition.
In the copying, "non-speech sound recognition" is changed to "no speech sound recognition".
Our two authors then return to copying from Cowling's 2004 thesis, continuing to reference 'comparison tables' that are not present in their 2014 paper. Almost the entire remainder of the section is directly from Cowling's thesis.
Plagiarised text
From looking at the comparison tables, it appears that some of these techniques, by their very nature, cannot be used for non-speech sound recognition. Any of the techniques that use sub word features will not be able to be used for non- speech sound identification. This is because environmental sounds lack the phonetic structure that speech does
There is no set “alphabet” that certain slices of non-speech sound can be split into, and therefore sub word features (and the related techniques) cannot be used. Due to the lack of an environmental sound alphabet, all of the Hidden Markov Model (HMM) based techniques that are shown in the table cannot be used. Since HMM techniques are the main techniques now used in speech and speaker recognition, this leaves only a few other techniques[6][8]. After discounting HMM, the remaining five techniques were tested for their ability to classify no speech sounds. This was done in two ways.
Original text
The comparison tables showed that some of these techniques, by their very nature, cannot be used for non-speech sound recognition. Any of the techniques that use subword features are not suitable for non-speech sound identification. This is because environmental sounds lack the phonetic structure that speech does. There is no set “alphabet” that certain slices of non-speech sound can be split into, and therefore subword features (and the related techniques) cannot be used (this is also noted in [ReyesEl03]).
Due to the lack of an environmental sound alphabet, the Hidden Markov Model (HMM) based techniques mentioned will be difficult to implement. However, this technique may be revisited in the future if necessary, and if a meaningful way of developing sound sub- units can be devised.
Frequently, when the 2014 paper's text includes references that appear out of order, it is trivial to perform a simple exact-text lookup and find that the section of text and its surrounding paragraphs were lifted with nearly no editing from an earlier study.
For example, backtracking to the first section, "I. Introduction", a study by "Vieira et al [5]" is cited. The fifth reference in its bibliography is not by Vieira et al, and thus this must be copied from a different study. Indeed, it is from the 2002 paper A System for Automatic Detection of Pathological Speech. The entire several-hundred-word introduction is lifted, once again verbatim with references intact, from this study.
This goes on. Rodrigues and Duraisamy's section explaining Mel-frequency cepstral coefficients is lifted verbatim from the introduction of the 2011 study Linear versus Mel Frequency Cepstral Coefficients for Speaker Recognition, citations and all.
Plagiarised text
Mel-frequency cepstral coefficients (MFCC) [1] have been dominantly used in speaker recognition as well as in speech recognition. This is counter intuitive to many researchers since speech recognition and speaker recognition seek different types of information from speech, namely, phonetic information for speech recognition and speaker information for speaker recognition. MFCC was first proposed for speech recognition and its mel-warped frequency scale is to mimic how human ears process sound. Its spectral resolution becomes lower as the frequency increases. Therefore, the information in the higher frequency region is down-sampled by the mel scale. However, based on theory in speech production [2][3], speaker characteristics associated with the structure of the vocal tract, particularly the vocal tract length, are reflected more in the high frequency region of speech.
Original text
Mel-frequency cepstral coefficients (MFCC) [1] have been dominantly used in speaker recognition as well as in speech recognition. This is counterintuitive to many researchers since speech recognition and speaker recognition seek different types of information from speech, namely, phonetic information for speech recognition and speaker information for speaker recognition. MFCC was first proposed for speech recognition and its mel-warped frequency scale is to mimic how human ears process sound. Its spectral resolution becomes lower as the frequency increases. Therefore, the information in the higher frequency region is down-sampled by the mel scale. However, based on theory in speech production [2][3], speaker characteristics associated with the structure of the vocal tract, particularly the vocal tract length, are reflected more in the high frequency region of speech.
Its Hidden Markov Model explanation just copies the 2014 Wikipedia entry on the concept.
Plagiarised text
The Hidden Markov Model have a wide range of application in temporal pattern recognition such as speech, handwriting, gesture recognition, part-of-speech tagging, musical score following, partial discharges and bioinformatics.
A Hidden Markov Model can be considered a generalization of a mixture model where the hidden variables (or latent variables), which control the mixture component to be selected for each observation, are related through a Markov process rather than independent of each other.
Original text
Hidden Markov models are especially known for their application in temporal pattern recognition such as speech, handwriting, gesture recognition, part-of-speech tagging, musical score following, partial discharges and bioinformatics.
A hidden Markov model can be considered a generalization of a mixture model where the hidden variables (or latent variables), which control the mixture component to be selected for each observation, are related through a Markov process rather than independent of each other.
The vast majority of Analysis of Speech Algorithms in Disease affected Voice Patterns is blatantly plagiarised.
Given the scale of the plagiarism on display, and the fact that only two people worked on this paper (which contains very little original data or research and next to no original writing), it would have been very difficult for either author (again, Dr. Paul Rodrigues and Dr. Karunanidhi Duraisamy) to be unaware of this misconduct.
It is with some disgust that I discovered this was not the only time the two authors collaborated on a paper nominally concerning Parkinson's disease.
In addition to the 2014 paper Analysis of Speech Algorithms in Disease affected Voice Patterns (covered above), Rodrigues and Duraisamy collaborated on four further papers.
- 2017: Diagnosis of Disease through Voice Recordings using Artificial Neural Networks
- 2018: Diagnosis of Parkinson's Disease Using Fuzzy Height
- 2018/19: A Fuzzy Rule-Based Diagnosis of Parkinson's Disease
- 2019: Fuzzy Inference System for analyzing the Parkinson's Disease Severity
All of these papers appear to contain egregious and blatant plagiarism.
Diagnosis of Disease through Voice Recordings using Artificial Neural Networks
The abstract of this paper contains a copied definition of neurodegeneration ("The process of impairment of brain cells is called neurodegeneration."), which originated on a website, Neurology Solutions, on a page explaining Parkinson's to medical patients.
The introduction plagiarises content that, while it has been copied to various Parkinson's explainer websites, first appeared at least as early as 2016, pre-dating this 2017 paper.
Plagiarised text
Parkinson’s Disease is a neurodegenerative brain disorder. A person’s brain slowly stops producing a neurotransmitter called dopamine. The less secretion of dopamine leads to less control to regulate their movements, body and emotions. The brains cells neuron produces the dopamine.
Original text
Parkinson’s Disease (PD) is a neurodegenerative brain disorder. A person’s brain slowly stops producing a neurotransmitter called dopamine. As dopamine decreases, a person has less and less ability to regulate their movements, body and emotions.
It then returns to copying from the prior explainer website, Neurology Solutions:
Plagiarised text
These neurons concentrate in a particular region of brain called the substantia nigra. Dopamine is a chemical carries information from the substantia nigra to other parts of the brain to control movements of a human body. When 60 to 80 % of dopamine cells got damaged, the symptoms of Parkinson’s Disease appears. This process of impairment of brain cells is called neurodegeneration [1].
Original text
Parkinson’s disease impacts nerve cells in a part of the brain called the substantia nigra. In PD, certain nerve cells called alpha-synuclein proteins break down. These are the neurons that produce dopamine, a chemical messenger in the brain. Dopamine acts as a transmitter for signals from the brain to other parts of the body and is essential for movement[...]
The process of impairment of brain cells is called neurodegeneration. When approximately 60 to 80 percent of the dopamine-producing cells are damaged, the motor symptoms of Parkinson’s disease appear
Much of the remainder of the paper's introduction is plagiarized from the 2014 paper Speech Analysis for Diagnosis of Parkinson’s Disease Using Genetic Algorithm and Support Vector Machine, including (once again, with citations) its explanation of the discovery of the disease and its explanation of voice-driven identification as a non-invasive early detection method.
Oh, and it opens its next section, "Materials", by copying that study's explanation of the Parkinson's speech dataset it uses to test its detection systems.
The authors then paste in half of a mangled table outlining the dataset's attributes, followed by an explainer for the table's contents cribbed directly (without attribution) from the dataset's documentation.
The paper cribs its explanation of neurons and neural networks from a 2009 paper. It takes its overview of the history of Parkinson's in India ("PD has been known in India since ancient days and the powder of Mucuna Pruriens seeds was used for its treatment.") directly from a 2003 paper.
The explanation of the trainlm function is taken near-verbatim from, of all places, the MATLAB Help Center.
While the experimental results appear to be original work, the bulk of the paper is plagiarised.
Diagnosis of Parkinson's Disease Using Fuzzy Height
In a change of pace, much of this paper is not directly copied from another source. It does, however, begin by uncritically drawing from a chapter of the 1986 book Movement Disorders. This chapter, Role of 8-Type Monoamine Oxidase Inhibition in the Treatment of Parkinson's Disease, is the source of the typo "Shaking Pulsy", which almost exclusively appears in badly written papers on Parkinson's disease that draw from this chapter.
Ill-considered text
PD occurs due to the neurodegenerative disorder in brain. James Parkinson deducted this symptom and named as “Shaking Pulsy” during 1817.
Original text
Parkinson's disease (PD), first described in 1817 as paralysis agitans by James Parkinson in his "Essay on the Shaking Pulsy," is characterized by tremor, bradykinesia, rigidity, and a postural defect.
A great deal of the remainder of the introduction draws heavily from now-defunct pages on parkinson.org, but perhaps not to a degree that indicates plagiarism.
However, significant parts of the following section, Materials, are badly copied from the MATLAB Help Center. Again.
Plagiarised text
Fuzzy Inference is a process of mapping the input to the output using fuzzy logic. The fuzzy infer ence involves Membership Functions, Logical Operations and If- Then Rules. Fuzzy rules are a collection of linguistic statements that describe how the FIS should make a decision regarding classifying an input or controlling an output.
Original text
Fuzzy inference is the process of formulating the mapping from a given input to an output using fuzzy logic. The mapping then provides a basis from which decisions can be made, or patterns discerned. The process of fuzzy inference involves all the pieces that are described in Membership Functions, Logical Operations, and If-Then Rules.
The following explainer of Mamdani-type inference appears to copy (again, verbatim) from a few sources that pre-date the paper. The sentence "Mamdani-type inference expects the output membership functions to be fuzzy sets" appears to be from a chapter in the 2013 book Metaheuristic Applications in Structures and Infrastructures, or at least the Science Direct explainer on the topic.
The following sentence, "The process of Mamdani system involves fuzzification of input variables, rule evaluation, aggregation and defuzzification" appears to be lifted fairly directly from one of any number of PowerPoint slideshows on the topic that have been on the Internet since 2013, all using identical phrasing to describe this process.
Standing alone, this paper is not particularly problematic. It definitely violates academic ethical norms by not citing where it draws much of its information from, and directly copies from some sources, but the bulk of the paper is a fairly minimal statistical analysis of the same dataset as in the authors' prior paper (Oxford’s Parkinsons Disease Detection Dataset).
A Fuzzy Rule-Based Diagnosis of Parkinson’s Disease
While I do not have access to this paper, it appears to be a relatively short (nine-page) conference paper.
However, the fact that its abstract appears to substantially consist of sentences directly copied from the authors' two prior papers on Parkinson's is not encouraging.
Fuzzy Inference System for analyzing the Parkinson's Disease Severity
The introduction to this paper includes paraphrased information from the authors' prior papers on the topic.
The methodology section opens with a sentence that, in a previous paper, was plagiarised from the MATLAB Help Centre: "Fuzzy Inference is a process of formulating the mapping from given input(s) to output(s) using fuzzy logic."
It then takes (and mangles) an explanation of fuzzy inference from a 2016 paper, Adaptive Neuro Fuzzy Interference System.
Plagiarised text
A FIS consists of three main components, fuzzy rules with “If-then”, as a function of the fuzzy set membership; and reasoning, fuzzy inference techniques from basic rules to get the output. Figure 1 displays the block structure of the FIS.
Original text
A FIS was built on the three main components, namely basic rules, where it consists of the selection of fuzzy logic rules “If-Then;” as a function of the fuzzy set membership; and reasoning fuzzy inference techniques from basic rules to get the output. Figure 2.5 shows the detailed structure of the FIS
The authors then plagiarise their explanation of the various types of FIS from the same paper.
Plagiarised text
There are several types of FIS, namely Takagi–Sugeno, Mamdani, and Tsukamoto. A FIS of Mamdani are most commonly widely used methodology.
Original text
There are several types of FIS, namely Takagi–Sugeno, Mamdani, and Tsukamoto (Cheng et al. 2005). A FIS of Takagi–Sugeno model was found to be widely used in the application of ANFIS method.
Notably, the copied-and-reworked section accidentally includes the first two words of the second sentence in the original text ('A FIS").
That sentence and the following sentence are also plagiarised, the detection of which is made easier by the inclusion of a distinct citation style copied from the source... which is, once again, MATLAB documentation.
Plagiarised text
A FIS of Mamdani are most commonly widely used methodology. Mamdani’s Fuzzy Inference method was the first control systems built using fuzzy set theory. It was proposed by Ebrahim Mamdani [Mam75]
Original text
Mamdani’s fuzzy inference method is the most commonly seen fuzzy methodology. Mamdani’s method was among the first control systems built using fuzzy set theory. It was proposed in 1975 by Ebrahim Mamdani [Mam75] as an attempt to control a steam engine and boiler combination...
Needless to say, copying content from MATLAB documentation is not acceptable academic practice.
Conclusions
Doubtless there are other parts of these papers that were plagiarised, but these are the sections that were most prominent.
The fact that these two authors, Dr. Paul Rodrigues and Dr. Karunanidhi Duraisamy, evidence the same pattern of plagiarism across all of their collaborations indicates that this is a deliberate behaviour.
The literature on Parkinson's disease is seeded with plenty of low-quality work, but these four articles (and likely the conference paper I do not have access to) are particularly shameless, appearing to represent a pattern of deliberate portfolio-padding by two academics who do not, as a matter of course, work with Parkinson's disease or even medicine.
The audacity of a geoscientist and software engineer to dip into a completely unrelated field to their core competences and clutter up healthcare research with their valueless slop to pad out their academia numbers, lazily stealing work from other authors in the process, is absolutely staggering.
It goes without saying that this pattern of academic dishonesty casts doubt over the experimental results in these papers, as well as the authors' work in other academic domains.