Challenges in obtaining medical data

Machine learning models rely on the training data provided. The algorithms learn from the examples available. It generally requires a lot of training data to develop models that are reliable and generalisable. The following are some of the challenges associated with getting workable data for medical projects :

• Unavailability of data : While it is comparatively easier to get training data for some applications of machine learning, that is not the case for medical data. As an example, ImageNet ^[1] , one of the largest available image datasets, which is also the basis of image classification benchmarking as a result of the ILSVRC challenges, has 14,197,122 available images ^[2] . Similarly, for studying text data, corpus could be generated from Wikipedia, Twitter and publicly available news articles and texts. But getting medical records of this magnitude is challenging. Gaining access to real world data, including EHR comes with its own set of issues.

• Lack of digitization in hospitals : There is a continued push towards making hospitals digital and using electronic health records (EHRs), with the main reasons in its favour being :

  • transparency

  • less medical errors due to illegibility or loss of written records

  • convenience to medical professionals with providing real time availability of old patient records^[6][5]

As per the American Hospital Association survey from 2009^[3] :

On the basis of responses from 63.1% of hospitals surveyed, only 1.5% of U.S. hospitals have a comprehensive electronic-records system (i.e., present in all clin- ical units), and an additional 7.6% have a basic system (i.e., present in at least one clinical unit). Computerized provider-order entry for medications has been implemented in only 17% of hospitals. Larger hospitals, those located in urban areas, and teaching hospitals were more likely to have electronic-records systems. Respondents cited capital requirements and high maintenance costs as the primary barriers to implementation…..

The above numbers have significantly improved in the US after the HITECH Act, but there are still issues as cited by Colligan et. al. ^[4] which leave hospitals dissatisfied with the EHR adoption and uses. It is also to be noted that while USA has been largely successful in increasing EHR usage among hospitals, so is not the case for most other countries. Many hospitals around the world still do not use EHR extensively and a lot of work in hospitals still happens on paper with little or no digitisation, with costs of setting up a digital system and amount of time and money needed to be spent in training the employees being a major barrier.

• Privacy : Even when hospitals do have electronic health data, it cannot be made publicly available or handed over to researchers easily. Proper licensing systems and guidelines need to be in place. It needs to be anonymised by the hospitals to ensure complete anonymity and privacy of patients ^[7]. HIPAA and recently GDPR have put regulations in place regarding handling of personally identifiable information which need to be respected, often leading to additional layer of complexity before preparing data to be made available outside hospitals.

• Challenges in data merging : When comparing and/or merging medical data from different sources, we may encounter following issues :

  • Language differences : If sources of data are from countries/ hospitals with different local languages, the terminologies may be different and it could take some time and data processing to ensure that correct values are being mapped to correct feature names.

  • Convention differences : As an example, while in one place, it may be usual to save body and weight separately, in other case, saving BMI and weight may be the norm. Relevant decoupling and variable generation may be required.

  • Differences in physical units : Different hospitals may follow different physical units. For instance, while Europe mainly follows the metric system, USA follows the United States Customary System (USCS) where in place of kilograms, pounds may be used while recording weights for the patients. Similarly for temperature, either Celsius or Fahrenheit may be used.

  • Different standardisations : Even when using standardisation conven- tions, a situation may arise that some data could be from an older time, or the hospital record system has not yet updated to a more recent standard. For instance, MIMIC-III dataset from the USA follows the ICD-9 ^[8] sys- tem, while ICCA database from Uniklinik Aachen follows the ICD-10 sys- tem ^{[10] [9]}. Even within the same ICD-x system, different revisions may be followed, and before continuing with data merging or any modeling, these need to be standardised.

  • Different privacy standards : Different hospitals may come under different privacy laws, meaning some data/features may be intentionally withheld from the provided dataset which renders that feature unusable, even if it is available in other datasets.

• Bias : Like all scientific studies, one needs to ensure that the data made available to a model is representative of the population over which it will be deployed and is not biased. [More on this topic in another post.]

References :

[1] Deng et. al ImageNet: A Large-Scale Hierarchical Image Database. CVPR, 2009

[2] http://image-net.org/about-stats

[3] Jha et.al Use of Electronic Health Records in U.S. Hospitals, The New England Journal of medicine, 2009

[4] Colligan et. al. American Medical Association Sources of physician satisfaction and dissatisfaction and review of administrative tasks in ambulatory practice: A qualitative analysis of physician and staff interviews.

[5] Linda T. Kohn, Janet M. Corrigan, and Molla S. Donaldson To Err Is Human: Building a Safer Health System (1999).Committee on Quality of Health Care in America, Institute of Medicine, National Academy Press, Washington, D.C.

[6] HIMSS Electronic Health Record Committee EHR Definition, Attributes and Essential Requirements Version 1.0

[7] https://www.legislation.gov.au/Details/C2012A00063

[8] International Classification of Diseases,Ninth Revision (ICD-9) https://www.cdc.gov/nchs/icd/icd9.htm

[9] ICD-10 versions on the WHO website https://www.who.int/classifications/icd/icdonlineversions/en/

[10] ICD-10 on CDC website https://www.cdc.gov/nchs/icd/icd10cm.htm

Tweet Share &source=https://kumarishikha.github.io" target="_blank"> Share -->