Lecture 3 Notes: Deep Dive into Clinical Data

6.S897/HST.956 Machine Learning for Healthcare Lecture 3: Deep Dive into Clinical Data 1 Understanding Clinical Data To start understanding Clinical Data, we begin with an example from the distribution of heart rates from the MIMIC-III database [EWJJPS+ 16] (as recorded in Careview), shown in Figure 1. This data involves around 600,000 admissions over a period of 12 years. Figure 1: The distribution of heart rates in the Careview medical system. Unusually for biological data like heart rate, the data is bimodal (we’d typically expect a distribution closer to normal). Why might this be? As it turns out, the hospital that provides this data switched care systems, from Careview to Metavision, and the old and new systems don’t record data in exactly the same way. A comparison of the two systems can be seen in Figure 2. Figure 2: A comparison of the heart rate data in the two different systems. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 1 The data from the Metavision looks closer to normal and is not bimodal, so it looks much more like we would expect. Further investigation reveals two major ways the first system differes from the second. First, under Careview, natal intensive care unit data was added to overall data set, while that data was not included in Metavision. Second, everyone over the age of 90 in the Carevision was listed as 300 years old upon their first visit to the system, in order to protect their identities in compliance with HIPPA regulations. These data are shown in Figure 3. Figure 3: Heart Rate vs Age (Careview). Once these differences are accounted for, the data actually look quite similar, which one can see in Figure 4. The lesson here is ”be careful with data.” There are many strange issues with how it’s collected and stored that can be confusing without additional information. Figure 4: Heart Rate vs Age for adults. 2 Types of Data There are many types of health care data we can use, including: • Demographics – This includes data like age, sex, race, etc. • Vital signs – These data are basically measurements a nurse would take during a regular check up, like weight, height, blood pressure, etc. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 2 • Medications – These data cover over-the-counter drugs one takes, as well as illegal drugs and alcohol. This is an area that patients could lie about, particularly when illegal drugs are involved. However, with lab results one can get a more accurate picture of the substances a person uses, in a process called ”medication reconciliation.” • Lab test results – Components of different bodily fluids, including blood, stool, urine, etc. • Pathology – This involves qualitative and quantitative examinations of any body tissues, including cell-level measurements such as cell-surface antigens. A rule of thumb is that if something is taken out of you during surgery, it’s probably going to Pathology. • Microbiology – This involves growing organisms, typically from cultures, to test their sensitivity to various antibiotics, at various dilutions, etc. • Notes – There notes at the end of medical reports, which can be quite long, and contain information like the kind of drugs a patient will be prescribed, whether there was a referring physician who sent the patient in, whether the patient has been advised to seek a specialist, whether the patient will receive in-home care, etc. • Billing – All the information about what was billed by the hospital. There can be a large amount of information here, because hospitals in general will want to bill for as much as they justifiably can. Includes ICD9/10 codes, procedure codes, etc. • Administrative data – This involves which service you’re on. An example of where this comes up would be that you need cardiology intensive care, but that service is full, and instead you get a bed in pulmonary intensive care. You would still be listed as getting cardiology service, even though your bed is in pulmonary. • Imaging data – this includes x-rays, ultrasounds, etc. • Quantified self data – Data that come from wearable devices, including steps walked, elevation change, heart rate, diet, blood sugar, etc. 2.1 Example Chart An example of the kinds of charts we might deal with is in Figure 5, which is a chart going over the care of a person in an intensive care unit from their admittance to their death. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 3 MIMIC-III, a freely accessible critical care database. Johnson AEW, Pollard TJ, Shen L, Lehman L, Feng M, Ghassemi M, Moody B, Szolovits P, Celi LA, and Mark RG. Scientific Data (2016). DOI: 10.1038/sdata.2016.35. Available at: http://www.nature.com/ articles/sdata201635 Figure 5: An example medical chart. The top row of the chart covers the amount the patient wants doctors to try to keep them alive if something goes wrong at the time in intensive care. ”Full code” means that they want every effort to be made to keep them alive, while ”Comfort measures” means that they want doctors to allow them to die if their condition worsens. In the chart, when the patient is admitted their code status is”full code,” but as time progresses they change to ”comfort measures.” The second section measures the physical capabilities of the patient, such as their ability to speak, their motor control, and their eye movements. While at the beginning of their admission they are in full bodily control, their condition deteriorates over their time in intensive care. The third section covers fluid measurements over the course of the visit, such as platelet count. The fourth section covers various medications that were administered to the patient over the course of their admission, including the doses involved. Finally, a chart covers various vital measurement over the course of the visit, such as heart rate and O2 saturation. 2.2 Demographic Comparison As part of an exploratory analysis, one can plot how demographic variables relate to each other to try to better understand subpopulations of the patients. For example, you could plot age of admission segmented by admission type (elective, emergency, or urgent), as seen in Figure 6. In this case, the distribution of age does not change very much depending on their admission type. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 4 Figure 6: Age of admission, by admission type. As another example, in Figure 7 we plot age segmented by insurance type. There is a clear change in age distribution here, as self paying customers skew younger, and most people switch to Medicare after age 65. Figure 7: Age of admission, by insurance type. More examples of these plots can be found in the lecture slides. One can also investigate how mortality is influenced by demographic information. In Figure 8 we have the results of generalized linear model trained on the health care data to predict mortality, with significant variables indicated by stars. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 5 Figure 8: Age of admission, by insurance type. Age being significant makes sense, as the older someone is the more likely they are to die. The statistically significant ethnicity variables are those that indicate ethnicity information is missing for some reason, which is difficult to explain – the information is missing frequently enough that it is unlikely it’s missing because people die before being able to identify their ethnicity. The remaining significant variables are knowing English or Spanish, which indicates being able to communicate with doctors more easily leads to lower mortality. 3 Different Medical Standards There are a wide variety of medical standards, for everything from prescriptions given to procedures and more. For example, in Figure 9, we see two patients with the same diagnosis given very different treatments. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 6 Figure 9: Two different treatments for the same diagnosis. One can also identify different medical standards by looking at the most common prescriptions in the database, as seen in Figure 9. For example, there are two different rows containing D5W, one with an NDC code and one without, along with several other examples of prescriptions without their NDC codes. Thus, even though ways of recording prescriptions like NDC codes are standardized across the US, the ways hospitals report what they prescribe are not necessarily standardized. Figure 10: Most common prescriptions. 3.1 Different Medication Coding Systems There are a large number of medication coding systems, including: • National Drug Code (NDC) – A 10 number identification code for a drug where the first four numbers identify who produced it, the next four the form of the drug, and the last two the number of doses. This coding system has the difficulty that it has run out of numbers for both the drug producers and the form of the drug, and attempts to expand it have not been applied systematically. • MedDRA – An identification system created by the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use. It isn’t compatible with NDC codes. • CPT Codes – The codes in the range 90281- 99607 give a variety of medicine codes. • 2019 Healthcare Common Procedure Coding System (HCPCS) – Used by Medicare and Medicaid patients. • Commercial Coding Systems – While often redundant, Medi-Span and First Data Bank also have created medication codes. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 7 3.2 Different Procedure Coding Systems There are also a variety of different codes that identify medical procedures patients receive, including: • ICD 9/10 procedure codes • CPT codes • MV codes CPT codes in particular are highly detailed – example CPT categories can be seen in Figure 11. Note that each category has multiple codes associated with it. Figure 11: CPT code categories. 4 Lab Reports The most commonly used coding system for reporting lab results is Logical Observation Identifiers Names and Codes (LOINC). It has a hierarchical structure where every lab test has a different code, but similar results are connected by categories. An example of a lab report can be seen in Figure 12. Note that the times of the tests have been de-identified, and some qualitative values like atypical lymphocytes do not have flags. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 8 Figure 12: Lab results for a patient. 5 Chart Events Chart Events capture a variety of vital sign features. Figure 13 contains some examples of such events. Some, like heart rate, appear twice. This could be an indication of two systems being combined to get these codes, which is something to watch out for. Figure 13: Chart events for a patient. There are also charts that capture patient outputs, like urine or stool samples. An example can be seen in Figure 14. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 9 Figure 14: Outputs for a patient. There are also tables for patient inputs, which include medications provided to them. Two examples of this type of table, one for CareVue and one for MetaVision), can be seen in Figure 15 and Figure 16 respectively. Figure 15: Inputs for a patient (CareVue). 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 10 Figure 16: Inputs for a patient (MetaVision). 6 Using Medical Process Measures to Make Predictions The required reading for this class, Biases in electronic health record data due to processes within the healthcare system [AKW18], showed that for many lab results, process measures of the data (such as the time a lab result was taken) are more important than actual values in predicting outcomes. While these results cannot be replicated exactly with the MIMIC III database, there are some proxies we can use to try to get similar results, such as white blood cell (WBC) counts. Looking at the fractions of abnormal white blood cell counts per hour, for example, matches the paper’s findings that tests taken in the early morning such as 4:00 am are connected to a person being unhealthy. The graph of this relationship can be seen in Figure 17. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 11 Figure 17: Proportion of abnormal WBC measurements per hour. One can also build a regression model to predict mortality from number of WBC measurements and number abnormal WBC measurements per hour. This can be found in Figure 18. Figure 18: Results for using regression to predict mortality from number of WBC measurements and number abnormal WBC measurements per hour. While there are some significant hours here, the fact that only some hours are significant and not others is not what one would expect based on [AKW18]. For example, if 8:00 am were a significant time to get a lab test done, one would think 7:00 am and 9:00 am would also be significant times, but in this regression that is not the case. Thus, it seems like there is some noise in the data causing the times to appear insignificant. We can use the MIMIC data to confirm that lab result values do vary by time of day, as the tables in Figure 19 show. This is one example of many tables in the lecture slides showing how lab tests results change 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 12 over the day depending on the time it is performed. While there are several possible explanations for this, such as diurnal human body changes or changes in care throughout the day, these results do align with the results from [AKW18] that the time a test is taken provides valuable information alongside the test results. Figure 19: Mean lab result values plotted against time for several lab tests. 7 Clinical Notes in MIMIC There are a wide variety of types of clinical notes – counts for the number of clinical notes taken by type of professional or visit in the data set can be found in Figure 20. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 13 Figure 20: Different counts of the number of clinical notes in the MIMIC database by profession or reason for visit. These notes can be very long. The empirical counts of the lengths of these notes taken by type of profession can be found in Figure 21 – average counts of 1000 words or more are not uncommon. Figure 21: The distribution of the lengths of the clinical notes by profession or reason for visit. An example nursing note can be found in the lecture slides. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 14 8 Data Encoding Standards OHDSI is the standard data encoding method. It is often used with Fast Healthcare Interoperability Resources (FHIR), which allows hospitals to share healthcare information electronically. The goal of FHIR is to provide the minimum amount of information a doctor needs to know to start treating a patient. Figure 22 shows an example of the form of healthcare information shared – various applications make the form easier for humans to parse. © source unknown. Figure 22: The distribution of the lengths of the clinical notes by profession or reason for visit. 9 Resources for Various Terminologies All of the terminology standards, such as LOINC, ICD9/10, etc., are gathered in the UMLS Metathesaurus at https://uts.nlm.nih.gov/home.html. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 15 10 Key Takeaways • ”Know your data” • Harmonising all the different types of data is difficult and time consuming • For some areas, standards don’t exist at all References [AKW18] Denis Agniel, Isaac S Kohane, and Griffin M Weber. Biases in electronic health record data due to processes within the healthcare system: retrospective observational study. BMJ, 361, 2018. [EWJJPS+ 16] Alistair Edward William Johnson, Tom Joseph Pollard, Lu Shen, Li-wei Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Edward Moody, Peter Szolovits, Leo Anthony G. Celi, and Roger G. Mark. Mimic-iii, a freely accessible critical care database. Scientific Data, 3:160035, 05 2016. 6.S897/HST.956 Machine Learning for Healthcare — Lec3 — 16