Modern Information Technology for Cancer Research: What’s in IT for Me? An Overview of Technologies and Approaches

Fighting cancer is the challenge of a lifetime for patients, their physicians and researchers in oncology. Hence, accepting all help available seems only logical. And where would one expect help in an era of digital disruption [1] if not in information technology (IT)? However, despite big promises to improve patient care and reduce errors or work load [2], using data and learnings from basic and clinical cancer research remains a challenge [3]. Even beyond scientific content, using IT tools today requires technical expertise and infrastructure, which is complex and costly to set up and operate. In this article, we give you a few impressions of scenarios employing technologies to support and improve cancer research or therapy. Following a prototypical process of research, we take you from solutions to explore a field by data, generate hypotheses and identify or share existing data within research networks. We advance to gathering and analyzing genomic data, opening up new data sources with patient engagement or machine learning in imaging, touch on techniques for analytics and visualization and end with support for experimental or augmented therapies. We believe such a journey can neither be exhaustive nor representative, as there is no registry for IT tools or specific key words to search for. However, we hope to give you – the oncologist or scientist in oncol ogy – a few ideas on how to get in the driver’s seat on the road to better outcomes.

Both oncologic treatment and research are for the most part still considered a data-poor environment [4], as most data is not systematically (and electronically) gathered, structured or integrated into IT systems. There are small areas that involve large data sets (e.g., omics data or radiology data), but data on processes, results or external knowledge still requires a lot of manual work or is not recorded at all.

Other sciences (e.g., meteorology and astrophysics) already have infrastructures in place to collect and share data because they invested in them [4]. To transform biomedical research – and cancer research in particular – into a data-rich environment, it will take a strong commitment to empower people and infrastructure. Many of the necessary building blocks already exist or are currently being researched – as is outlined below. However, there are many other areas that need more attention, especially by physicians driving innovation for the sake of their patients. One big area to focus on should be electronic health records (EHRs). They need to go beyond representations possible on paper and provide structured information in order to become useful for research [5]. A few good examples are already out there, as some companies develop systems for oncologists to use for both care and research [6].

Existing IT solutions will not cure patients by themselves; they need to be employed by caretakers in all medical professions. Many advances in medical IT research are yet to be introduced into routine care. Some apprehensions about investing in health IT are that more reliance on IT could introduce error [7], divert attention to other patients [6] or simply interfere with workflows [8]. However, if implemented wisely, it can open a window of opportunity for faster research, more insights and therapy options and ultimately a better outcome for patients.

Starting our journey towards new insights, we want to explore the field of oncology to identify areas to improve cancer therapy or form hypotheses for pathogenesis. Data on current diagnostics, therapies and their outcomes could provide a great start. Information on patients collected during diagnosis or treatment is usually stored in an EHR. Even though EHRs are not designed for research, they represent an unprecedented source of patient information for research if properly augmented [4]. Naturally, it has long been a goal to make this data available for research and many attempts have already been made in many projects [9-11]. Unfortunately, poor data quality and quantity remain a challenge [12, 13]. It is possible to extract clinical information from EHRs using natural language processing [14]. However, as of this writing, extracted information can be produced at scale, but is rarely as useful as structured information [15]. Hence, improving documentation of structured data and thereby improving data quality remains the best option to gain usable data for research. However, quantity can still remain a challenge. But it can be overcome by sharing data, for example, within a research network.

Research sites and institutions will always have limited resources like staff, facilities, study participants or samples. Therefore, it seems logical to cooperate with other researchers to form a research network. Within these networks, knowledge and samples can be exchanged, but data exchange seems even more crucial. Whether to increase the sample size in a study or to compare different studies performed, more data promises to improve results dramatically. Building such networks is possible, as has been proven by the German Cancer Consortium [16]. However, integrating data documented at independent institutions is a massive undertaking, especially if it was originally documented for treatment purposes [17]. Since the data to be shared is found in various documentation systems, a multitude of technical interfaces have to be developed [18] (syntactic interoperability). And as staff at different sites have slightly different conventions to document findings, mapping the various meanings to a common data set is far from trivial (semantic interoperability). Data privacy is also an issue, as the data relates directly to cancer patients. Well-established IT solutions for pseudonymization are freely available [19] but have to be integrated into daily processes by each institution. For example, obtaining the informed patient consent required for most forms of data processing [20] (EU-GDPR) occupies the physician’s already scarce time. A decentralized data storage concept employing centralized metadata and well-defined processes for data access can help to comply with regulation and still yield respectable results.

Though controversial from a data privacy standpoint [21], the advantages of cloud computing are enticing: virtually unlimited resources for storing data, analyzing and visualizing it on demand plus instant access for everyone authorized [22]. The Seven Bridges Cancer Genomics Cloud is an example for such a cloud to analyze genomic databases without the need for a local infrastructure to hold the large data sets [23]. By providing access to these data sets from different sources, the cloud can enable machine learning or finding similar cases, as they (generally) rely on a large number of samples. IBM Watson, for instance, could identify drug candidates by sifting through massive amounts of data in pilot studies [24]. Beyond storing data, cloud computing can also execute algorithms supplied by researchers directly on the data stored [22, 23]. This enables other researchers to use the algorithms and gain insights from their output without the need for deep technical knowledge [25]. Lastly, by strictly versioning the data sets, the algorithms used and the infrastructure, the results of any given combination are reproducible [25]. Looking back on our scientific journey, we have now covered sharing and access to data. Next up is analysis – focusing on genomic data.

Genomic diagnostics promises to change cancer research and ultimately cancer therapy for good. So, it is likely that most cancer research will incorporate it in the future. Even today, therapy for cancer patients is guided by findings in their tumor genome [26]. However, the quantity and quality of proven causal relationships between genome changes and clinical implications are still low [26] compared to centuries of “conventional” medical knowledge. Phenotype information is needed to improve research on genome-related diagnostics and to create a learning healthcare system [27]. Unfortunately, obtaining both genomic and phenotype data from the same patients can be very difficult. Additionally, phenotypic data is often of low quality (i.e., incomplete or imprecise) or is not semantically comparable (i.e., the same condition is recorded in different ways). Ultimately, for patients to benefit from genetic diagnostics, a physician has to correctly interpret the results and draw conclusions [26]. This is by no means a small task and, besides costs, probably the biggest obstacle towards widespread adoption. Hence, we need more certainty in predictions based on genetic data and coupling it with large amounts of other patient data could help to fill that void.

We set out our journey to gather information on the outcome of cancer therapy, which is ultimately the condition of patients. Involving them in their care and opening up new data sources for research is next.

It takes innovative approaches in cancer therapy a long time to be accessible to physicians and patients. One underlying issue in cancer treatment research is the lack of trials, more specifically the lack of participation therein. Patient participation in trials is low compared to the number treated (in the US 2–3% of adults) [6]. Besides other concerns, like fear of placebo-treatment for instance, the effort for participants seems to be a major barrier. Well-designed informatics approaches could help with that. A good example can be drawn from cardiology: Data gathered from wrist-worn devices or from patient reports through their smartphones was included in a study in a large scale [28, 29]. Increasing patients’ engagement in their treatment can improve their experience and outcome [30]. It can also yield valuable data for cancer research.

mHealth – health information and services provided through mobile devices – can increase patient engagement; however, there are many challenges, either technical, legal or human to overcome [31]. Data on the quality of life for patients (as measured by standardized methods like the 36-Item Short Form Survey [32] or EuroQuol-5 Dimensions [33]) – especially those undergoing cancer treatment – could form a valuable basis to enhance treatment decisions [34]. First, results promise psycho-oncological benefits for patients with cancer [35]. Beyond informing patients, mobile devices also allow them to share information on their physical or psychological status and needs. Evaluating the outcome or current condition of patients beyond mortality has recently received more attention in oncology, especially by health payors or insurers [34]. This attention is highly warranted, since integrating patient-reported symptoms into clinical care could increase survival for patients with metastatic cancer [36]. Enabling patients to submit their information to their physicians or caretakers could also make it available for research. Getting information straight from the source like this can highly increase the quantity of data-points. For instance, a detailed pain history (with many data-points) can indicate an insufficient pain medication in comparison to a single value at a scheduled appointment. In the United Kingdom, patients report on their results and status to assess the quality of certain surgeries performed [34]. Altogether, patient-reported information could be very valuable for cancer research. But there are more sources we can tap, as we continue to derive data from images.

Image-based diagnostics currently focusses on machine learning based on deep convolutional neural networks (CNN) and radiomics. Radiomics [37] aims to extract quantitative, reproducible, structured, machine readable information out of medical images. Combining the 2 can have a significant impact: One example improves mammography screening by significantly reducing the false positives findings and therefore avoiding unnecessary invasive biopsies [38]. The system can automatically detect breast lesions on diffusion-weighted MR images. Another example is the improved stratification of Glioblastoma regarding overall survival and progression-free survival by adding radiomic subtyping to molecular, clinical or standard imaging parameters [39].

Other applications aim to assist in identifying image pattern in microscopy to reveal hints for cancer. A group at Google created an augmented reality application that uses machine learning to highlight suspicious areas in samples [40]. They trained a model to extract these areas based on sample areas previously identified as cancer. This example shows the huge potential that exists in combining computers, which are good at recognizing patterns consistently, and humans, who can check results for plausibility and reason on them. Ultimately, Radiomics opens up the treasures of data buried in medical images to research and analysis for both humans and machine learning algorithms.

So far, our journey through IT in cancer research took us to analyzing data and augmenting it with more data from innovative sources. Making sense of vast amounts of multiform data – a process described as analytical reasoning [41] – is now required to derive new insights. Out of all the findings in clinical information – patient-reported results, genetic variants, imaging data – which ones are significant? And how do they relate to studies, case reports, medication trials or cellular pathways? Research and good examples hint at what is possible with visualization [41-43]; however, there is no report of widespread use of these technologies in real-world environments [43]. Especially within a research context, visualizations are a chance to clean data sets from error-laden data or detect patterns more easily [44]. Also, by means of visualization, one can explore a data set to understand unexpected trends or hints, even without a sophisticated analysis [45]. In a research setting, this is very valuable for hypothesis generation or preliminary analysis (e.g., for feasibility studies). Remarkably, domain scientists (like cancer researchers) have trust in – and will hence likely benefit quickly from – newly introduced visual analytics tools when they are intuitive, transparent and controllable [46]. Their level of trust in visual tools is even higher than that in manual analysis methods when performing complex tasks [46]. Thus, visualization has benefits beyond efficiency gains and could provide researchers with better insights. As these should quickly find their way into therapy, we propose to cover therapy assistance next.

Personalized diagnostics based on genomic data is in high demand, since changes in genomic material can have huge effects [47-49]. Hence, Molecular Tumor Boards are established for biologists, pathologists and oncologists to collaborate in finding targeted therapies for patients, who do not benefit from standard treatment [50]. For the Molecular Tumor Boards at the National Center for Tumor Diseases [51] in Germany bioinformatics pipelines analyze patient and cancer cell transcriptome and exome, search for modifications and filter for those known to be susceptible for medication [52]. Using this information, physicians gather known publications, past or present trials, known gene relations and further knowledge on specific variations or mutated genes. Most tools are freely available web applications and do not require installation and are therefore easily accessible:

cBioPortal [53] provides an overview of gene variation and their domains. The public version accumulates data from currently 248 studies. Like with similar tools [54], one can search for one or more specific genes to obtain information on studies, gene relationships or clinical relevance. CIViC (Clinical Interpretations of Variants in Cancer: https://civicdb.org/home) also focusses on clinical relevance. One can query for gene-wide or specific variants and receive precise information on variant types and drugs in various phases, from in-vitro studies to results of phase 3 clinical trials.

Information about completed studies is often found in literature databases, such as PubMed. Queries can be optimized using MeSH (Medical Subject Heading) Thesaurus, which is used to index all articles in PubMed [55]. Embase (https://www.embase.com) by Elsevier is a pay-to-use database and has the advantage of intuitive search options like PICO, allowing the option of filtering by population, intervention, comparison and outcome.

As the number of studies focusing on molecular markers is increasing steadily, ClinicalTrials.gov (https://clinicaltrials.gov/) can list studies to enroll the patient in or learn from their – sometimes even preliminary – findings. This manual task of matching a patient to trials (finished or ongoing) is complex [56] but could be automated in the future [57].

Surgical data science aims to use all available data of the care process and integrate it to assist physicians and caregivers [58]. As machine learning is applied to solve specific problems, the shortage of curated data sets becomes a bottleneck. In minimally invasive surgery, for example, it is extremely challenging to capture all the variations (e.g., all medical devices applied in an intervention) that may occur in clinical practice. Furthermore, algorithms trained on data of one specific domain (e.g., specific hardware for a specific intervention) typically cannot be generalized well for new domains. Approaches proposed to overcome these challenges include the following:

1. Crowdsourcing: To spare resources of (medical) experts, annotation tasks may be outsourced to masses of untrained workers from an online community [59, 60]. The workers’ clickstreams (e.g., pattern of cursor movement, number of clicks/zooms) can be automatically analyzed [61] to estimate the quality of annotation. Also, hybrid human-crowd annotations concepts have been proposed for large-scale medical image annotation at low cost [62].

2. Self-supervised learning: While annotated medical data is rare, unlabeled medical data is much more easily accessible. Based on this observation, Ross et al. [63] investigated the hypothesis that masses of unlabeled video data can be used to learn a representation of the tar get domain that boosts the performance of state-of-the-art machine-learning algorithms. In other words, they showed that CNNs can even train themselves to recognize structures within video material, given that they were previously exposed to example images of the same domain.

At the end of our journey, many, if not all, of the presented scenarios, specialized systems or platforms provide information or answers for individual questions. Unfortunately, cancer research or treatment yields complicated questions – and answers – at high speed. Maintaining up-to-date information and processing new information are daunting tasks. The sheer volume of information to query and process is at some point destined to overwhelm a human’s capabilities – no matter how smart or educated. Hence, the challenge is not just in making information available but also in the capability of processing a vast amount of information and filter for actionable insights.

Combining some of the techniques shown, like data extraction, visualization or inclusion of genomic data could speed up hypothesis generation for research. However, in order to yield valid results, we need high data quality and good metadata – data describing data. Metadata can ensure that data sets – or rather their content – are valid and comparable. For instance, by enforcing the same vocabulary to describe certain conditions or provide a way to convert them.

Once supplied with good data and metadata, systems for presenting it have to be crafted in order to reduce the complexity. Since the types of questions and hence analyses differ between researchers, they each require their own system. Unfortunately, there is no “plug-and-play” system for research data available. Some portals combine data sets or offer different workflows, but they are not truly customizable (yet).

Still, you can take the next step, if you are willing to learn (and invest some time). Engage in improving EHRs and other documentation systems, so that entered data can be reused for scientific purposes! We believe that open data sharing and improving data quality can yield exceptional results and are first steps towards a “learning health care system” [3, 64]. But it will require commitment, investment and the willingness to cooperate of many researchers involved. As long as control over a specific data set yields a competitive advantage and scientific partners are not trusted, the system at large will not benefit as much as it should.

Despite all the obstacles outlined, cancer research is en route to a data rich environment, which could enable a much faster turn-around time. In common life science studies, data is collected to test one specific hypothesis and rarely reused. This process is very labor and time intensive. Preexisting data-pools and well-curated meta-data could enable testing a hypothesis in about a day’s time – or even less. Machine-learning approaches could speed this up even more or broaden the scope of analyses. We believe data-rich cancer research will empower researchers to spend their time more effectively. Basic research will always be necessary, but it could be guided more effectively. Data- and knowledge-driven clinical cancer research could help to bring innovation from the labs to the patients a lot faster.

Just like a person is much more than the sum of her parts, treating a patient with cancer needs a holistic view. Considerable progress has been made to understand individual processes and tools have been crafted to unlock previously unknown secrets. But putting the pieces together requires the oncologist, who sees her patient as a whole and is determined to find the best possible therapy by incorporating all that modern science has to offer. Modern IT tools can be a way to create and leverage scientific insights and bring them to the bedside, just as it can supply researchers with valuable data about real patients.

The authors declare that they have no conflicts of interest to disclose.