Evolving from data to knowledge: Using physiologic data to facilitate clinical decisions

Better data means better decision making, implying better care

The trend in data collection is that of evolution from episodic, ad-hoc, to more continuous collection of higher frequencies. This higher frequency data collection increases the likelihood that events will be “caught” — hence, better surveillance.

More continuously available information undergirds better care
More continuously available information undergirds better care

“Expert systems are the most common type of clinical decision support system technology in routine clinical use.” [1]

The types of data and uses for expert systems include:

  • Alerts & Reminders
  • Therapy Planning
  • Prescription
  • Data & Information Retrieval
  • Image Recognition & Interpretation

Common Clinical Decision Support System Myths [2]

  • Diagnosis is dominant decision-making issue in medicine: Not “what does patient have,” but, rather, “What should I do?”
  • Clinicians will use expert systems if they can be shown to function at the level of experts: “What do we know?”
  • Clinicians will use standalone tools, so long as they are integrated into clinical workflow.

What is the motivation for better Decision Support?

All across the specialty landscape the prediction for increased demand within the next 20 years is considerable. The motivation for more clinicians is clear. Yet, in addition to increasing the available pool of care providers, improving the capabilities of those in the field through providing better knowledge and decision making capability through access to the right information is key to improved care.

The growing demand and decreasing supply motivates the need for better decision support tools using the available data.
The growing demand and decreasing supply motivates the need for better decision support tools using the available information.

Display is Key to the Right Decisions.

The famous graphic by Charles Minard showing Napolean’s Russian Campaign, is a testament to the genius of his graphic design around 6 data elements: (1)army size, (2,3) position (x,y), (4) direction, (5) temperature, (6) time. Knowing how to display and integrate data into the workflow of the clinician is key to the usability as well as providing a valuable tool for informing on clinical decisions.

Data in 6 variables from Minard (1861) showing Napolean's Russian Campaign. Illustrated are: army size, (x,y) position, direction, temperature, time
Data in 6 variables from Minard (1861) showing Napolean’s Russian Campaign. Illustrated are: army size, (x,y) position, direction, temperature, time

The original presentation by John R. Zaleski is available for download here.

References

[1] Enrico Coiera, Guide to Health Informatics 2nd Edition, Chapter 25, October 2003.

[2] Edward Shortliffe, “Medical Thinking: What Should We Do?” Conference on Medical Thinking, University College of London, June 23rd, 2006.

Tracking an Object in Free-Fall Using the Discrete Kalman Filter

Modeling Ballistic Free-Fall

The white paper, which can be downloaded below, describes the use of the Kalman Filter in tracking the vertical position (altitude) of a spherical object. The purpose of this white paper is to illustrate the system dynamics, control, and tracking algorithm application to a system of known true dynamics. Gauss-Markov noise is applied to the measurements to simulate uncertainty in observing the true position of the object. The mathematics and theory can then be applied to other systems.

Update — May 4th, 2016: Thanks to an astute reader (Jeff Reimer) who discovered an error in my calculation for the cross-sectional area of a sphere. I used the wetted area instead of the cross-sectional area, which is an oversight from my fluid dynamics days over 30 years ago. I have uploaded the paper with correction to Equation 3 (Kalman Filtering of Ballistic Free Fall Object) here:

BallisticFreeFallTracking-update-1

Kalman Filtering System Dynamics

It is necessary first to establish a model of the system dynamics. This is a mathematical representation of the applied external forces that define the motion of the sphere as it descends. Knowing the system dynamics, the accuracy of the prediction of the object’s pathway is increased immensely, particularly if the object is accelerating at a known rate. If the object system dynamics are not known, it is still possible to track the object, with the system dynamics being derived from the observed motion. Yet, the predicted motion of the object and the accuracy with which its position and speed are known from measurement to measurement will not be as accurate and can deviate particularly if external forces cause unpredictable accelerations on the object as it descends. Consider the diagram of Figure 1, in which a sphere of mass, m, and radius, r, is falling under the influence of the two principal forces weight and drag .

Figure 1: Forces on a sphere in ballistic free-fall.
Figure 1: Forces on a sphere in ballistic free-fall.

The attached white paper is developed into a Java program to compute the track of the object as it falls. The purpose of this posting is to provide the capability to those interested in reproducing for academic reasons and for general education. The white paper will be used as the starting point for other posts I intend to write on the tracking of physiological processes, such as heart and lung dynamics.

Medical Device Integration: Growth, Trends & Challenges: An Interview with John Zaleski, PhD, CPHIMS

HIMSS Blog: Discussing Medical Device Integration Post Publication of John R. Zaleski’s Book “Connected Medical Devices…”

I was interviewed for the HIMSS Blog by the Editors of HIMSS Books and the interview in total is available for viewing at the following link:

Interview with John R. Zaleski: Medical Device Integration: Growth, Trends & Challenges

Why Medical Device Integration Now?

From my recent HIMSS Blog post:

More than half of U.S. hospitals and health systems are planning to purchase and implement a medical device integration (MDI) solution. This is quite a difference from, say, 5 years ago. There are a number of reasons motivating this. Partially, the maturing deployment of electronic health record systems; partially, the maturing of the complexity of integration that requires higher-frequency, higher accuracy, higher fidelity data, such as clinical decision support methods within electronic health record systems; partially, the motivation of Meaningful Use and needs for improvement in patient safety; partially the PP-ACA. Other specific motivations, such as the recognition that improved patient care management can be achieved through better, more accurate data. Furthermore, MDI is an essential element for achieving better patient safety.

Need for Higher Fidelity Data Drove Medical Device Integration Exposure

As a researcher with more than 20 years working with medical devices and as a product developer and inventor, the following trends and major milestones are, in my opinion, the recognition of the value of MDI, which occurred not too long after electronic health record systems became widespread, perhaps not quite 10 years ago; then, the motivating Federal guidelines surrounding Meaningful Use Stages 1 & 2; the PP-ACA also provided some motivation. But, beyond that, my opinion is that receipt of data into the electronic health record systems motivated new ideas about what to do with those data. This, in my opinion, is leading to “higher-level” use cases other than charting. For example, use of the data to assist in improved patient care management and clinical decision making. When I started in this field, I was a graduate student and needed to collect data on live patients whom I was studying to develop methods for weaning from post-operative mechanical ventilation. I was running a study on patients recovering from coronary artery bypass grafting surgery and was following “my” patients from surgery through to extubation from mechanical ventilation in surgical intensive care and general surgery. When I was conducting this study, it was years before the commercial electronic health record system was widely publicized. Furthermore, none of the medical devices with which I was working had any automated data collection capability that was exploited within the hospital system I was working. Hence, I had to write my own code and perform data collection on my own, right at the patient bedside. My purpose was in using the data to develop models of patient state; to better predict time-based changes in physiologic and respiratory parameters, and guided by patient demographics, intakes and outputs, and other information. Yet, what was really lacking was a way to collect this information using an automated, standardized approach. So, I got into the MDI field as it was a necessary utility to meet my ultimate needs: complete data for better clinical decision making.

Healthcare Analytics at INFORMS 2015: Kalman Filtering, Medical Device Integration Standards, & Medical Device Actionable Alerts

Three analytics presentations and session at INFORMS 2015 on prediction using Kalman filtering; medical device integration standards; and medical device alarms

This coming week, I will be giving three presentations and chairing a panel on the emerging role of health systems engineering, with impact on clinical informatics and analytics. This is a joint session between INFORMS and HAS.

This panel convenes on Tuesday, November 3rd from 16:30-18:00. The formal title of the session is “The emerging role of health systems engineering and its impact on clinical informatics and analytics.”

The three speaker presentations during this panel session will focus:

  1. J. Venella, DNP, RN, “How to make clinically actionable alarms”
  2. J. Zaleski, PhD, CPHIMS, “The Kalman filter and its application to real-time physiologic monitoring of high-acuity patients”
  3. S. Jha, “Predicting the effect of introducing walk in hours on staff workload at a pediatrics practice”

On Wednesday, November 4th, from 11:00-12:30, I will be giving the following presentation:

J. Zaleski, “Why are medical device connectivity standards so elusive?”

This presentation will focus on the history and current pathway toward standardized medical device integration and data communication. Medical devices still remain highly proprietary in terms of interoperability. Health Level Seven (HL7), as a healthcare information standard, only works when medical devices can export data in this common format. Gaps remain between the proprietary, manufacturer-specific language of many devices and the HL7 messaging format. Here we explore approaches for standardizing proprietary equipment around HL7 and related messaging languages and how lack of interoperability impacts patient care.

Finally, on Wednesday, November 4th, from 14:45-16:15, I will be giving the following presentation:

J. Zaleski, “Identifying patients at risk using fuzzy logic”. The use of “big data” for decision making has been a growing area of investigation and usage in healthcare enterprises. This paper shows how fuzzy rules can be used to operate on data obtained from the point of care to assist in clinical decision making, with application to real-time data collection in medical surgical units. This presentation will give an example of patient care management of medical device using data obtained from the patient at the bedside.

Systems Engineering in Critical Care Medicine: An Evolving Field

Systems Engineering

As we approach INFORMS 2015, one session theme is health systems engineering. Within the body of this session, the focus on examples of systems methods in medicine is exemplified by the various presentations.

Title slide from Symposium given at the University of Pennsylvania to Systems Engineering Graduate Students & Faculty. (c) 1999
Title slide from Symposium given at the University of Pennsylvania to Systems Engineering Graduate Students & Faculty. (c) 1999

This briefing presented results and research into postoperative critical care medicine. Specifically, weaning of patients who have undergone coronary artery bypass grafting. The parameters and automatically obtained data from bedside medical devices were recorded and analyzed to assess the trajectory of spontaneous breathing recovery over time. The systems engineering aspect involved bringing multiple variables together (integration) and generating a prospective model that was validated on a training set of data, then applied to a test set. The model’s skill in prediction demonstrated that the respiratory state of the patient (measured in terms of respiratory rate, minute volume) evolved from total dependency to near extubatory values with a prediction accuracy of approximately 1 hour (i.e., the ability to predict ahead the patient’s state by 1 hour).

Modeling and Simulation Key Aspects of Systems Engineering

The model demonstrated an application of systems engineering: specifically, the ability to integrate multiple data together and create a high-level model of respiratory state, guided by observations. Also demonstrated was the automated decision support mechanism of a manual process guided by a clinical protocol. In effect, this research demonstrated a control algorithm that provided a feedback response on maintaining mandatory support while responding to patient spontaneous capabilities. Achieving success with the controller is dependent upon having access to real-time clinical observations (both automated measurements from medical devices and clinical observations from staff). Laboratory data was also necessary as this provided the key as to when safe orders to begin weaning could occur. The analysis involved inferring trends and behavior from data; providing for visualization of data to promote visual projection to future state; using raw data to assist in guiding outcomes of weaning performance; and providing a means for linking medical data repositories to achieve access to complete records of data. This research was also conducted in the years prior to the global rollout of electronic health record systems. Thus, much of the data collection needed to be “cobbled” together from raw methods developed in-house.

Data collection were conducted under institutional IRB #570-0 involving live human subjects.

An Excel Method for Optimal Assignment of Resources Using the Hungarian Method

Optimal Assignment of Resources

Those working in operations research or analytics who are faced with the challenge of performing optimal assignment of resources (for example, assigning sales people to routes to minimize airfare cost), may have the need for the algorithm I am about to discuss. In general, the type of problem described is the assignment of n resources to n tasks such that at most only 1 task can be assigned to 1 resource such that the total cost of assigning the task to resource is minimized.

Optimal Assignment Cost Matrix
Optimal Assignment Cost Matrix

The Hungarian Method (a.k.a. Munkres algorithm) is a method that determines the optimal assignment of a given cost matrix. The cost matrix, as shown above, reflects the relative costs of assigning resources in column j to actors in row i. For small matrices (i.e., 2 or 3 assignees), the solution to determining the overall minimum cost can be computed relatively quickly by hand without the use of an automated algorithm. Yet, when considering larger assignments (e.g.: 4 or more rows and columns), the use of an automated method is necessitated as the computational cost grows as n! (i.e., n-factorial).

That is a big number.

An Excel Macro for Hungarian Method Optimal Assignment

The constraint that the cost matrix must be square (i.e., at MOST 1 column assigned to 1 row) is necessary for the algorithm. There are extensions to the Hungarian method that support rectangular assignment (i.e., selecting the best columns to assign to the rows when the cost matrix is not square). For our purposes, I will present only the nxn example.

The method I will provide to those who contact me directly. Have had a number of people ask me for other methods I have written in the past, and would prefer a direct reach-out versus simply posting on my site. I will share the screen shots of its operation, though.

Here is the cost matrix input--worksheet named "CostMatrix". Two buttons link to running the assignment and resetting the assignment, as shown.
Here is the cost matrix input–worksheet named “CostMatrix”. Two buttons link to running the assignment and resetting the assignment, as shown above.
Pressing the "Run" button will cause the algorithm to execute. The output will be written to an "Output" worksheet. Once the assignment is completed, the assigned entries are displayed in the cost matrix in mauve.
Pressing the “Run” button will cause the algorithm to execute. The output will be written to an “Output” worksheet. Once the assignment is completed, the assigned entries are displayed in the cost matrix in mauve.
Here is a view of the output worksheet. The row and corresponding optimally assigned column are displayed together with the total cost (or residual).
Here is a view of the output worksheet. The row and corresponding optimally assigned column are displayed together with the total cost (or residual).
Here is a sneak peak at the macro. Those interested can contact me here through the Web-Log for details.
Here is a sneak peak at the macro. Those interested can contact me here through the Web-Log for details.

This summarizes the optimal assignment method. It operates in Excel (all versions) and is quite lightweight.

Discrete Wavelet Transform Compression of Clinical Data to Support Rapid Storage and Retrieval

Discrete Wavelet Transform as a Tool for Data Analysis

As medical devices are increasingly required to provide for information from the point of care to enterprise electronic health record systems, more patient data collected at the point of care will become available for remote viewing and analysis. In this paper a discrete wavelet transform method is presented for automatically filtering and analyzing data obtained from high acuity bedside physiologic monitors using discrete wavelet transforms, the purpose being to reduce the total amount of data transmitted into the electronic health record, and to facilitate analysis of time-based trends in physiologic data.

Use of Discrete Wavelet Transform Compression

As information technology has been brought into the healthcare enterprise, much of the paper-based record is being supplanted by an electronic record, in which clinicians either record information manually or, in addition, automatically from clinical systems. The electronic medical record is maintained by the health care enterprise and follows the patient throughout all phases of diagnosis and treatment. Furthermore, this medical record is accessible to all authorized clinical personnel. An obvious benefit of this approach is that, unlike the paper record, the electronic medical record can be accessed from many different locations without physically retrieving the patient’s hard copy information from particular departments. Loss of information is nullified, and use of the electronic medical record establishes a standard approach for recording of patient information, so that each department must conform to specific standards in terms of the types and quality of information recorded on each patient. Also, with Web-browser-based medical record viewing, convenience in terms of viewing, together with the reduced delays associated with retrieving the paper recording ensure that clinicians can readily obtain patient information when required. In addition, two-way communication between the enterprise information system and clinical systems enable the error-free retrieval of patient demographic and administrative information (such as medical record number and insurance information) without adding further delay or introducing errors into the patient’s record within the departmental system.

One key difference between the legacy paper record and the electronic medical record is that the paper record remained an intimate device by which the attending nurse monitored and recorded status on the patient: it remained with the patient and the nurse until the patient left the unit. With the introduction of the electronic medical record, data transmitted to that record and viewable by authorized individuals outside of the unit can lack the context of the actual situation in interpreting patient flow sheet results. As society and healthcare move toward a completely automated and electronic medical record environment, it must be mindful of the fact that the introduction of new technologies must never impede quality healthcare [1].

In this paper the author describes a process for augmenting the basic transmission of information from the clinical environment to the medical record by capturing more detailed information that may not be captured during the course of standard flow sheet recording. Patient telemetry is normally recorded within the flow or assessment sheet on regular intervals. Typical intervals range from 15 minutes to an hour, depending on the particular acuity of the patient. However, bedside monitors typically can record very detailed information in fractions of a second. Most of this information is discarded, and much of it can be of no clinical value at these short intervals. However, a trade-off exists in terms of the size of the interval and the capturing of relatively important data from these bedside monitors: make the recording interval too large, and events of relatively short duration but high importance (such as heart rate spikes, or respiratory rate increases) will be missed and never recorded within the electronic medical record. On the other hand, make these cording intervals too small, and the health care enterprise, including the hospital computing network and the size of the medical record, will become cumbersome and filled with much useless information, possibly even rendering the system unusable. One approach to solving this problem is to provide the capability to record detailed information when necessary, but omit when not.

Use of Discrete Wavelet Transform as a Method for Data Compression

Performance and response time is a key metric when communicating raw clinical data from departmental to electronic health systems within a healthcare enterprise.  The objective of this paper was to present a discrete wavelet transform method that can reduce the overall data storage requirements as well as facilitate data analysis of time-varying signals without requiring large scale data mining of the raw information. The discrete wavelet transform (DWT) was selected as a possible filtering mechanism because the DWT preserves both spatial and temporal behavior of a raw data signal. This is a very important feature in the study of medical telemetry, because many processes are not stationary, making the application of traditional signal processing methods (such as Fourier transforms) inappropriate. The creation of a DWT Processing method that exists as an adjunct to the existing departmental information system imposes no additional software features on the existing telemetry system, and operates off of the existing clinical network. Furthermore, the benefits of using the DWT Processor as both a noise filter and as an automatic data filter are affirmed inasmuch as both stationary and non-stationary signals are processed appropriately using the DWT method: stationary signals can be represented by relatively few overall data points in the form of wavelet coefficients, whereas threshold filtering of non-stationary signal can provide accurate reconstruction of raw signals with even a factor of two fewer data points than the original signal. This benefits a potentially congested network and speeds recreation of the original signal by requiring fewer overall calculations to be performed by that enterprise information system.

A copy of the PDF version of the paper is included here:

Discrete Wavelet Transform.

[1]  Patricia Benner, “Beware of Technological Imperatives and Commercial Interests That Prevent Best Practices,” American Journal of Critical Care; 12(5): 469-471, 2003.

Read and Write Color Bitmaps Using Visual Basic inside of Microsoft Excel

Developers working with graphics frequently use applications that can read and transform image files. Whether they are transforming bitmaps into other file formats, or just reading the images in order to create new images from the pixel data, developers in this field regularly work with routines that access the base pixel data. Software methods written in C, C++ and Java are often used to read and write these files. The white paper attached below demonstrates the reading and writing of bitmap files using  Microsoft Excel using Visual Basic. A benefit of this approach is the Visual Basic interpreter inside the Macro editor is used, making the method extremely powerful and simple to code and use — no formal compilers or knowledge of C, C++ or Java is required to implement this set of Visual Basic methods.

The  simple Visual Basic method to read and write bitmaps runs in an Excel spreadsheet, which reads a .bmp file. This method opens, reads, and extracts the pixels of a 24-bit color bitmap, in accordance with a user-specified color. This method then writes the extracted pixels to a new bitmap containing only the selected color specified by the user. This straightforward method requires no compiler.

The original Microsoft Visual Basic bitmap reader/writer method was written in 2002: Read and Write Bitmaps Using Excel.

MD+DI Article: Medical Dvice Integration (MDI): A World of Possibilities

A new article published in Medical Device and Diagnostic Industry (MD+DI) discusses how middleware used for medical device integration (MDI) as part of the Medical Device Data System (MDDS) FDA clearance can migrate medical devices from electronic health record charting to active patient monitoring, for the purpose of intervention and guidance:

Minimally, MDDS middleware needs to be able to retrieve episodic data from a medical device and translate it to a standard format.”