A bibliometric review of peripartum cardiomyopathy compared to other cardiomyopathies using artificial intelligence and machine learning

Abstract

As developments in artificial intelligence and machine learning become more widespread in healthcare, their potential to transform clinical outcomes also increases. Peripartum cardiomyopathy is a rare and poorly-characterised condition that presents as heart failure in the last trimester prior to delivery or within 5-6 months postpartum. The lack of a definitive understanding of the molecular causes and clinical progress of this condition suggests that bibliometrics will be well-suited to creating new insights into this serious clinical problem. We examine similarities and differences between peripartum and its closely related familial dilated cardiomyopathy and idiopathic dilated cardiomyopathy. Using PubMed as the source of bibliometric data, we apply artificial intelligence-supported natural language processing to compare extracted data and genes association with these cardiomyopathies. Gene data were enhanced with additional metadata from third-party datasets and then analysed for their impact and specificity for peripartum cardiomyopathy. Artificial intelligence identified 14 genes that distinguished peripartum from both dilated and familial dilated cardiomyopathy. They are as follows: CTSD, RLN2, MMP23B*, SLC17A5, ST2*, PTHLH, CFH*, CFI, GPT, MR1, Rln1, SRI, STAT5A* and THBD. We then used the Human Protein Atlas website that uses affinity-purified rabbit polyclonal antibodies to identify genes that are expressed at the protein level (bold), or as RNA transcripts (*) in healthy human left ventricles. Additional analysis focussed on the full set of peripartum genes on linkage and specificity to cardiomyopathy yielded a different set of thirteen genes (bold font indicates those expressed in cardiomyocytes: PRL, RLN2, PLN, ST2, CTSD, F2, ACE, STAT3, TTN, SPP1, LGALS3, miR-146a, GNB3, SRI). This type of analysis can highlight new avenues for research, aimed at improving genomics-driven peripartum cardiomyopathy diagnosis as well as potential pathological and clinical sub-classification. We expect that this will allow for future improvements in identification, treatment and management of this condition. The first step in the application of these bibliometric-based artificial intelligence methods is to understand the current knowledge, and it is the aim of this paper to show how this might be achieved.

Keywords: Artificial intelligence; Bibliometrics; Genomics; Machine learning; Peripartum cardiomyopathy.

Fig. 1

The human pregnancy cycle with PPCM incidence where the columns represent the incidence; orange illustrates the circulating levels of prolactin (Prl) and soluble fms-like tyrosine kinase-1 (sFlt-1) during a normal pregnancy (Arany and Elkayam 2016)

Fig. 2

Schematic summary of the key cardiac contractile proteins (adapted from Peng et al. 2014)

Fig. 3

Review framework showing the process steps used in this review

Fig. 4

The number of articles per year obtained from on Bibliometrics in PubMed indicates that this is a rapidly growing field of research

Fig. 5

The timeline of journal articles published in PubMed (note the log scale)

Fig. 6

Venn diagram showing the overlap of CM, DCM, FDCM, and PPCM journal articles by PMID in the literature

Fig. 7

Venn diagram showing the overlap of CM, DCM, FDCM and PPCM genes in the literature

Fig. 8

Overview of the 23Strands’ Biblioengine V2.2

Fig. 9

The global locations of primary author affiliations in genomics research shown across PPCM, FDCM and DCM, the size of bubbles represent counts of publications

Fig. 10

Scientific Evolutionary Pathway of PPCM and genomics, note how the topic evolution extends up and down from the first topic of pregnancy complications (Cardiovascular (2005))

Fig. 11

A study of author co-occurrence revealed clusters around Sliwa (purple), Hilfiker-Kleiner (grey), Elkayam (green), Ricke-Hoch (orange) and Ikeda in the top left (light orange). The size of the bubbles matches the number of articles published. The grid numbers have been used to disambiguate the authors

Fig. 12

Country co-occurrence; this clearly shows collaborative research across Europe, Africa and the United States and a number of islands of research including Japan

Fig. 13

Gene plus sub-disease co-occurrence shows the multitude of co-morbidities identified in the research title and abstracts and the co-occurrence with the major genes

Fig. 14

Gene plus chemical co-occurrence highlighting two clusters of inflammation and cardiogenic shock

Fig. 15

Centrality analysis of PPCM Genes with relationship to cardiomyopathy, cardiac contractile proteins highlighted in red. Centrality is measured as a composite of the influence (degree centrality), topological distance (closeness centrality) and the shortest path (betweenness centrality) of the gene nodes relating to the topic of “cardiomyopathy” in the scientific literature on PPCM (Freeman et al. 1979)

Fig. 16

A vasculo-hormonal hypothesis of the pathophysiology of PPCM (adapted from Bello and Arany 2015), during the peripartum period there is an increased secretion of hormones from the pituitary (e.g. prolactin (PRL)) and placenta (e.g. sFLT1). This change is believed to contribute to the underlying cardiac dysfunction in PPCM