1. How do doctors think?

How do doctors think? In other words, how do they use their medical training to interpret a patient's signs and symptoms to develop a.) a diagnosis and b.) a treatment plan?
1. Medical decision making involves various degrees of uncertainty. Some decisions are based upon deductive reasoning. Others are based on empirical knowledge or experiences. Physicians can have a prior probability of what the likelihood of a disease will be and a pretest probability after they get more information. Physicians can use their clinical intuition to examine the etiology (cause) of the symptom. Physicians can also order tests. The more a test reduces uncertainty the more useful said test is. One model for physician cognition breaks is down into three categories.
  1. Representativeness
  2. Availability
  3. Anchoring and Adjustment
2. When reading testing results physicians have to take into account the confusion matrix. The confusion matrix is a 2x2 matrix that has 4 categories: True Positives, False Positives, True Negatives, and False Negatives. By examining the ratios of these statistics called sensitivity and specificity, doctors can make informed decisions based on testing.

2. Where might clinical decision support be useful?

One major failure point of physician decision making is assessing probability. Clinical decision support can help physicians estimate probabilities in a few ways. First, clinical decision support can take into account more factors than a physician could with some models taking in hundreds of covariates. Secondly, clinical decision support can utilize the deluge of biomedical data and is not restricted to a particular doctors patient load. Third, clinical decision support is objective and, for example, it won't weigh a more recent example than a past one.
1. Another major failure point in physicians is bias. While algorithms can have bias, good one’s can be trained not to. Three biases that can occur when examining test results are spectrum bias, test-referral bias, and test-interpretation bias. One very common other bias is confirmation bias.While algorithms aren’t absent of biases, they are less likely to succumb to biases like confirmation bias.

3. Clinical decision support system

Clinical decision support (CDS) will provide stakeholders, such as clinicians, staff, patients, or other individuals with knowledge and person-speciﬁc information, intelligently ﬁltered or presented at appropriate times, to enhance health and health care. Systems that provide CDS come in three primary varieties: 1) retrieve relevant online documents; 2) provide patient-specific alerts; and 3) data visualization. The steps involved in CDS include prevention, diagnosis, prognosis, treatment, and evaluation. CDS aims to promote patient safety by allowing the health care system to share data with stakeholders.

One interesting situation in medicine called "hallway consult (Links to an external site.)" refers to the informal consolation that the treating physician seeks informal information or advice about patient care or the answer to an academic question from a colleague. Although that advice from colleagues might be partial and result in malpractice, the health practitioner who made the final decision and performed the clinical practice should responsible for the wrong result. Due to the limited scope of knowledge and time, it is impossible for a human to expert in the multidisciplinary field with several domains. Therefore, computer-based programs that analyze data within EHRs and related online documents to provide reminders to assist health care providers in implementing evidence-based clinical guidelines at the point of care for a specific patient can be taken as the CDS.

4. Doctor's orders

Computerized provider order entry (CPOE) systems aim to replace the paper-based ordering system in the hospital and health care system. Mostly, CPOE involves the process of providers entering and sending treatment instructions – including medication, laboratory, and radiology orders. CPOE enables users to enter the orders, maintain medication records, and review changes. CPOE systems can, when correctly configured, markedly increase efficiency and improve patient safety and patient care by offering safety alerts.

The mnemonic "ADC VanDIMLS (Links to an external site.)" helps medical students to memorize the hospital admission orders and care, which aims to improve quality of care and outcomes for hospitalized elderly. Remembering the mnemonic "ADC VanDIMLS" assist medical students in identifying the issues with the highest priority that must be addressed during hospitalization to provide the optimal outcome.

CPOE offers safety alerts when an unsafe order is entered, as well as clinical decision support to guide caregivers to less expensive alternatives or to choices that better fit established hospital protocols. However, facilities need to recognize that currently available CPOE systems require a tremendous amount of time and effort to be spent in customization before their safety and clinical support features can be effectively implemented. What's more, even after they've been customized, the systems may still allow certain unsafe orders to be entered. Thus, CPOE systems are not currently a quick or natural remedy for medical errors. In conclusion, CPOE can solely support decision making instead of making the decision. ("Computerized provider order entry systems," 2001)

5. Success and Failure of CPOE/CDS - in research and application

In the White Paper by AMIA Board of Directors in 2007 (Osheroff, et al., 2007), a structure to maintain a stable CDS by three fundamental components was proposed. The whole framework is shown as below.

Figure 1. Three pillars for enhanced health and health care through CDS

In the framework, three pillars are:

Best Knowledge Available When Needed
High Adoption and Effective Use
Continuous Improvement of Knowledge and CDS Methods

Each of the pillar were accompanied by two strategic objectives. For details, refer to https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2213467/.

In theory, CPOE and CDS systems, invented in 1990s, should promote better and more efficient health care with fewer error. During the practice of implementing, people realized that user interface is a very important factor to make whatever fancy ideas come to real usage. A lot of sad stories about CPOE/CDS would be connected how stupid a mistake may seem like, such as clicking the wrong diagnosis codes just one item below the correct one, reckless of typing in an extra 0 in dosage, assigning wrong drugs due to 10 patients’ records opened up at one single time, etc.

Another key factor is identifying the right workflow in different clinical settings. Different specialties have different focus on clinical practices, leading to their unique needs for CPOE. A CPOE for pharmacists may not work well for orthopedic surgeons. An example of the workflow of CPOE in ICU settings was studied by Cheng, et al (Cheng, et al., 2003), as following, with key conceptual components of coordination redundancy, computational interface, and work location:

Figure 2. Order entry workflow

A better user interface will definitely help. Besides, those systems still have a much room to improve, technology-wise. Some errors could be prevented by showing up a critical warning when physician trying to assign a drug with crazy high dose, or fetal drug interaction. This would naturally lead to the old topic of warning fatigue, common to all those systems. The trade-off here becomes more important from place to place, specialty to specialty.

7. Related Literature

7.1 Literature review for NLP for clinical decision support

Demner-Fushman, D., Chapman, W. W., & McDonald, C. J. (2009). What can natural language processing do for clinical decision support? J Biomed Inform, 42(5), 760-772. doi:10.1016/j.jbi.2009.08.007

Natural language processing (NLP) is instrumental in using free-text information to drive CDS, representing clinical knowledge and CDS interventions in standardized formats, and leveraging clinical narrative. Active NLP CDS includes alerting, monitoring, coding, and reminding. The currently existing systems roughly fall into two categories: general-purpose clinical NLP architectures (increasingly publicly available), and specialized systems developed for specific tasks.

Figure 3. An integrated self-governed multi-task NLP-CDS system (Demner-Fushman, Chapman, & McDonald, 2009)

General-purpose clinical NLP systems

MedLEE

MedLEE is an NLP system that extracts information from clinical narratives and presents this information in structured form using a controlled vocabulary. MedLEE uses a lexicon to map terms into semantic classes and a semantic grammar to generate formal. representation of sentences. Have no idea why this is not currently used in clinical settings.

Text Analytics

The Text Analytics architecture developed in collaboration between the Mayo Clinic and IBM is using Unstructured Information Management Architecture (UIMA) to identify clinically relevant entities in clinical notes.

2. Specialized clinical NLP systems

Clinical events monitoring: detection and prevention of adverse events.
Processing radiology reports
Processing emergency department reports
Processing pathology reports
Processing a mixture of clinical note types

Screen Shot 2020-04-24 at 9.52.05 PM.png

7.2 Clinical Decision Support Systems: A survey of NLP-based Approaches from Unstructured Data

Reyes-Ortiz, J. A., Gonzalez-Beltran, B. A., & Gallardo-Lopez, L. (2015). Clinical Decision Support Systems: A Survey of NLP-Based Approaches from Unstructured Data. 2015 26th International Workshop on Database and Expert Systems Applications (DEXA), 163–167. https://doi.org/10.1109/DEXA.2015.47

This paper provides a survey of Natural Language Processing techniques for clinical decision support. It’s a computational challenge to use natural language processing techniques on biomedical data. The challenges depend on whether the data is structured, semi-structured, or unstructured. This paper focuses on unstructured data which is estimated to be the hardest. In this case, unstructured data is used to mean free-text. NLP is used for information extraction (IE) which is extracting structured information from unstructured data as well. This article outlines the medical domain, nlp task, patient-type, clinical decision support task, and health outcome for several medical domains. The main nlp approaches were syntactic, semantic, and statistical. All in all, this article concludes that unstructured data is being leveraged for clinical decision support in a number of specialties and languages.

7.3 Transition from CPOE to Paper-based Systems?

Griffon, N., Schuers, M., Joulakian, M., Bubenheim, M., Leroy, J. P., & Darmoni, S. J. (2017). Physician satisfaction with transition from CPOE to paper-based prescription. International journal of medical informatics, 103, 42-48.

Griffon et al did a quite interesting research on physician satisfaction when transiting from CPOE to paper-based prescription, a practice that seems to go in an opposite way of current trends. The study happened in Rouen University Hospital in Rouen, France, where CPOE was first introduced in 2012 and announced a failure three years later in 2015, and then they rolled back to paper-based order entry “system” (PBOE). The survey, covering 51 respondents using both systems, showed increases in satisfaction (OR=3.74), usability (OR=4.00), reliability (OR=8.54), and communication with nurses (OR=14.27). Some positive views about paper-based system include the ability to view multiple things at the same time (such as biology results, prescription, care plan, etc), stability of never being out of order. Main struggles with CPOE are: crashes and dependency on network and devices, delays in transmission and loss of prescription, bad suggestions in complex cases, and disruption of communication process between nurses and physicians.

8. Reflection Blog

Reflection blog for Part 3: CPOE/CDS: click me!

References:

Computerized provider order entry systems. (2001). Health Devices, 30(9-10), 323-359.

Demner-Fushman, D., Chapman, W. W., & McDonald, C. J. (2009). What can natural language processing do for clinical decision support? J Biomed Inform, 42(5), 760-772. doi:10.1016/j.jbi.2009.08.007

Osheroff, J. A., Teich, J. M., Middleton, B., Steen, E. B., Wright, A., & Detmer, D. E. (2007). A roadmap for national action on clinical decision support. Journal of the American medical informatics association, 14(2), 141-145.

Cheng, C. H., Goldstein, M. K., Geller, E., & Levitt, R. E. (2003). The effects of CPOE on ICU workflow: an observational study. In AMIA Annual Symposium Proceedings (Vol. 2003, p. 150). American Medical Informatics Association.