Trends in Alzheimer's Disease Research Based upon Machine Learning Analysis of PubMed Abstracts.

About 29.8 million people worldwide had been diagnosed with Alzheimer's disease (AD) in 2015, and the number is projected to triple by 2050. In 2018, AD was the fifth leading cause of death in Americans with 65 years of age or older, but the progress of AD drug research is very limited. It is helpful to identify the key factors and research trends of AD for guiding further more effective studies. We proposed a framework named as LDAP, which combined the latent Dirichlet allocation model and affinity propagation algorithm to extract research topics from 95,876 AD-related papers published from 2007 to 2016. Trends and hotspots analyses were performed on LDAP results. We found that the focus points of AD research for the past 10 years include 15 diseases, 15 amino acids, peptides, and proteins, 9 enzymes and coenzymes, 7 hormones, 7 carbohydrates, 5 lipids, 2 organophosphonates, 18 chemicals, 11 compounds, 13 symptoms, and 20 phenomena. Our LDAP framework allowed us to trace the evolution of research trends and the most popular areas of interest (hotspots) on disease, protein, symptom, and phenomena. Meanwhile, 556 AD related-genes were identified, which are enriched in 12 KEGG pathways including the AD pathway and nitrogen metabolism pathway. Our results are freely available at https://www.keaml.cn/Alzheimer. © The author(s).

Introduction

Between 2010 and 2015, the number of Alzheimer's disease (AD) cases has more than tripled, and the prevalence of AD is still increasing 1,2. 2015 report estimated that 29.8 million people had been diagnosed with AD 3,4. As a chronic neurodegenerative disease, AD is the cause of 60% to 70% of dementia cases and in 2015, the G8 nations endorsed AD as a major societal concern when it was reported as the cause of about 1.9 million deaths 5-9. Moreover, the number of global AD cases is projected to more than triple by 2050 2,10,11. In 2018, AD became the sixth leading cause of death in the United States and the fifth leading cause of death in Americans with 65 years of age or older 12. It is predicted that by 2050, the number of deaths caused by AD will account for 43% of all deaths among the elderly in the United States 4.

AD is a type of brain disease that begins primarily at the age of 65 or older, although 4% to 5% of AD cases are early-onset 13. The early symptom in AD is difficult in remembering recent events, also called short-term memory loss 6,14. Symptoms arise because part of the brain's nerve cells (neurons) that are involved in thinking, learning, and memory (cognitive function) have been damaged or destroyed15. As the disease advances, symptoms may include problems with language, disorientation, and some pathological signs and symptoms such as atrophy and paralysis. As an AD sufferer's physical condition declines, much of his or her basic bodily functions are gradually lost, and this will ultimately lead to death 16. Although the health situation may vary after diagnosis, the life expectancy is 3 to 9 years on average 17,18.

Many journals, conferences, patents, and associations are dedicated to the study of AD. For example, Alzheimer's Association releases an annual report, which describes the public health impact of AD, including incidence and prevalence, mortality and morbidity, use and costs of care, and the overall impact on caregivers and society 19. There are nearly 0.6 million authorized patents about AD. With the fast development of biotechnology and medicine, a huge amount of biomedical literature and data have been published. For example, until February 2019, 142,785 papers have been published on AD in PubMed. From these research, we can find the research hotspots and trends of AD.

For the curation research on AD, the whole world is struggling to find more effective ways to treat the disease, delay its onset and prevent its development. However, this progress is not satisfactory. For example, it is reported that researchers had spent more than 100 billion dollars to find effective drugs to cure AD in clinical treatments in past years. But the results were disappointed as in the failure of the antibody, solanezumab, which held the promise of improving cognition but failed trial tests 20. Alzheimer's disease is currently incurable, but treatments for symptoms are available. Although current Alzheimer's treatments cannot stop the disease progressing, they can temporarily slow the deterioration of dementia symptoms and improve the quality of life 14,21. In 2018, it is reported that more than 16 million family members and other unpaid caregivers provided approximately 18.5 billion hours of care for people with Alzheimer's or other dementia. The value of this care is close to $234 billion 15. AD is considered as one of the most financially costly diseases in developed countries 22,23.

For clinicians or biological researchers, rapid and effective acquisition of cutting-edge information on research advances from tens of thousands publications is a huge challenge 24. However, because of the complexity of the multi-disciplinary study, there is no research to summarize these findings to keep pace with the research trends of AD using machine learning and natural language processing models. To address this problem, we adopted machine learning methods to automatically summarize the trends and hotspots of AD based on PubMed abstracts. For this purpose, we collected 10 years of abstracts on AD research from PubMed. Then we proposed a framework named as LDAP which combined latent Dirichlet allocation model and the affinity propagation algorithm to extract research topics and hotspots. By incorporating the Medical Subject Headings term categories, we studied the trends includes medical entity changes in diseases, chemicals and drugs, symptoms and phenomena, and gene pathway based on the hotspots for the past 10 years. There may be other trends changed about AD, but in this study, we only focus on the changes in the areas mentioned above.

Methods

LDA topic model

Latent Dirichlet allocation (LDA) is a topic model algorithm based on probability proposed by Blei et al. in 2003 25. It is an unsupervised machine learning technique and can be used to identify potentially hidden topic information in large-scale document sets or corpora 26. LDA is a document theme generation model, which also known as a three-layer Bayesian probability model. It assumes that each word is extracted from a hidden theme behind it, and it contains three layers of words, topics, and documents 27,28. LDA assembles a collection of documents, where D={d}, m∈{1, …, M}. The distribution of topic k over vocabulary is denoted as Φ={ϕ}, k∈{1, …, K}, and the distribution of the m-th document over all K topics is denoted as Θ={θ}, m∈{1, …, M}. A topic is a distribution of terms over a vocabulary. It allows each document to be described as a distribution over topics, which can be expressed as:

where w represents a word, d represents a document, and t is the topic. It can be expanded as follows:

where, for document m, the distribution of document over topics θ and the distribution of topics over vocabulary Φ are sampled from priors α and β, respectively. Then, the topic assignment z for each word is generated from θ, and the accurate words w are generated according to their respective topic assignment z and the distribution of topics over the vocabulary Φ.

To get the topic words of each year, we used Gibbs sampling to estimate the LDA model. Gibbs sampling is a Markov chain Monte Carlo (MCMC) algorithm for obtaining a sequence of observations which are approximated from a specified multivariate probability distribution, when direct sampling is difficult.

Affinity Propagation model

Affinity propagation (AP) is a clustering algorithm based on the concept of “message passing” between data points. AP finds "exemplars", which are members of the input set representing clusters 29. AP has been used in many fields, such as image processing 30, text clustering 31, and gene detecting 32.The description of an AP algorithm is as follows: Given a sample set X = {x1, x2, ..., x}, and there is no hypothesis of inherent structure between data. Let S be a matrix that depicts the similarity between points, s(i, j) > s(i, k) if and only if the similarity between x and x is greater than that of x and x.The similarity is represented by the reciprocal of cosine similarity corresponding to (3):

The AP algorithm alternates between two messages passing steps to update two matrices:

(a) The responsibility matrix R: r(i, k) describes the extent to which sample k is suitable for the clustering center of sample i, and it represents the messages from i to k.

(b) The availability matrix A: a(i, k) describes the selection sample i choosing sample k as the degree of suitability for the cluster center, and it represents the messages from k to i.

The matrix R will be constantly updated according to (4).

The matrix A will be continually updated according to (5) and (6).

After copious iterations, the final clustering result can be obtained when the two matrices converge. To avoid numerical oscillations, a damping factor λ∈(0,1) is introduced when updating the two matrices, as described in (7) and (8):

Proposed framework

This paper presents an LDAP framework, which combines LDA model and AP algorithm to extract research topics. Our proposed framework is depicted in Figure 1. It consists of data processing, methods, and trends analyses. The data from PubMed are downloaded and pre-processed with operations of word segmentation, lemmatization, and stop words removal techniques. After data acquisition and preprocessing, the data is fed to the topics model LDA. The evaluation index, perplexity, in the LDA model can cause some variation in the number of topics. When the number is too small, the model may not reach convergence and achieve the optimal result; meanwhile, the model will lose the ability of capturing topic-diversity and this can cause highly used words like ad, disease, alzheimer, and dementia to get high scores in each topic. Therefore, we generated topics using LDA and then clustered them with AP. With the introduction of MeSH, David, and KEGG, the results can be represented from different aspects, such as proteins, diseases, and pathways.

Figure 1

Proposed framework. This framework consists of data processing, methods, and trends analyses. Data downloaded from PubMed was pre-processed and fed into an LDA model. AP algorithm was used to cluster the topics generated from LDA. Trend analyses were done on these topic words.

Experiments and Results

Data set and preprocessing

Because an abstract can provide a concise and accurate description of the important content of a paper, necessary information can be obtained from abstracts without reading the full text. We took “Alzheimer's disease” as the entry term to search papers in PubMed and obtained 95,876 papers from 2007 to 2016. Each file is a semi-structured XML document and contains various tags, such as