Decision Analysis in Healthcare

Objective Risk Analysis Can Reduce
Unnecessary Security/Privacy Fears

 

Farrokh Alemi, Ph.D.

Vikas Arya

Version of June 16th 2005

This research was supported in parts by the National Capital Region Critical Infrastructure Project (NCR-CIP), a multi-university consortium managed by George Mason University, under grant #03-TU-03 by the U.S. Department of Homeland Security’s Urban Area Security Initiative, and grant #2003CKWX0199 by the U.S. Department of Justice’s Community Oriented Policing Services Program. The views expressed are those of the authors, and do not necessarily reflect those of the Dept. of Homeland Security or the Dept. of Justice.

Summary

We provide an objective and quantitative method for analyzing risk for unauthorized disclosure of private patient information.   The current method, experts generate scenarios of how security or privacy violations might occur and organizations protect themselves against these risks.  The current approach leaves healthcare organizations at mercy of vivid imagination of their consultants.  In contrast, we propose a method that relies on objective data regarding actual occurrences of privacy violations.  The proposed method relies on a National Incidence Database of Unauthorized Disclosures (NIDUD) to identify a small set of risk factors that have led to unauthorized disclosure.  Using this set of factors, Health Care Organizations complete a survey and report the frequency of the hazards within their organization.  Bayesian probability models are used to compute the overall probability of unauthorized disclosure from survey responses.   This paper provides a tutorial on how to do objective risk analysis within health care organizations as well as a test of the concept.  We show the application of the approach to four health care institutions. 

Introduction

Health Insurance Portability and Accountability Act (HIPAA) and its subsequent final rules[1] require all healthcare providers, their suppliers, contractors and business affiliates to conduct a comprehensive analysis of the risks for unauthorized disclosure. This regulation has caused organizations to rush to analyze risks.  Many consultants have emerged with various methods for assessing organization’s vulnerability.  In this paper we propose a new method for risk assessment.

Our approach is based on the National Incidence Database of Unauthorized Disclosures (NIDUD); in which health care organizations voluntarily report their sentinel HIPAA-related events.  The database is used to focus the risk assessment on specific hazards that have occurred in at least one other health care organization.  In contrast, today health care organizations conduct risk assessments based on imagined risks that might mislead them to protect against events that will never occur.  They may waste precious security funds.  Even worst than a one time waste is the prospect that when another consultant, with more active imagination and a more vivid assessment tool, shows up; then the health care organizations is catapulted to invest more -- chasing elusive and esoteric security targets.  Since imagination is limitless, there is no end to how much should be spent on security and which vulnerability is more important.  Like a child, the organization ends up fighting imaginary foes.  Risk assessment instead of helping the organizations focus on high-value targets, misleads them to pursue irrelevant targets.  When analysis is based on real vulnerabilities/threats that have led to privacy violations, organization can focus on probable risks and rationally prioritize and limit investment in security controls. 

Objective analysis of security/privacy risks seems implausible on several grounds.  First, privacy and security violations remain rare and therefore assessments of these probabilities are difficult, if not impossible.  Second, objective risk assessment is not practical because of extensive data needed.  Third, the assessment seems to focus on addressing problems that have arisen in the past and ignores the inventiveness of efforts to break the system. This paper addresses all three criticisms.  We show that it is possible to assess rare probabilities from historical patterns, even when the event of interest has happened rarely or even only once.  We calculate probability of the event from time-to-the-event (e.g. time-to-computer-theft inside an organization).  In this fashion, very small probabilities can be assessed.  We show that such estimates can be used to calculate risk of unauthorized disclosure. 

Are the procedures practical and can they be accomplished in a short time frame by real organizations?  Surprisingly, a focus on objective risks reduces as opposed to increase the data collection burden significantly to a small set of risk factors.  This makes the objective risk analysis far less time consuming and more practical than a typical comprehensive risk analysis.  We show, by way of four examples, how real organizations can assess overall risk and priorities quickly and based on objective data. 

The last concern with objective risk analysis focused on preparing against inventiveness of people who want to break the system.  The proposed objective analysis does not prepare against all possibilities and leaves the organization partially vulnerable.  But no one can plan against unknown risks; neither our proposed approach nor the existing state of the art of doing everything possible protect us against the unknown.  No analysis, no matter how comprehensive, can anticipate unknown risks.

Objective risk analysis is useful in creating market-based incentives for improving privacy of patient health information. If accurate methods for assessing probability of unauthorized disclosures existed, then insurers can offer products for coverage of HIPAA liabilities. Then, the premiums for HIPAA insurance will create market incentives for organizations to improve their privacy and security. Organizations with a poor track record will face higher insurance premiums and those with adequate security safeguards will face lower premiums. This paper provides procedures for conducting probabilistic risk rating from experiences of various organizations; thus it provides a first step in offering HIPAA insurance.

History

The methodology for probabilistic risk analysis has evolved over the years through its application in the aerospace, nuclear power, and the chemical industries.[2] The Health care industry can build its risk assessment procedures based on progress made in these industries. A key feature of risk assessment procedures in these industries has been a cumulative database of hazards and scenarios from accidents or near-accidents. For example, following Challenger disaster, the National Aeronautical and Space Administration had to revise its procedures for risk assessment to include scenarios previously not considered in the predictive models and to correct and revise the probabilities of other types of failures.[3]  Similarly, after the Three Mile Island nuclear accident, risk assessment procedures had to be revised and the probabilities associated with human error had to be significantly increased.[4] Experiences of other industries highlight the need to take an evolutionary approach to risk analysis, to share information across organizations and to increase the knowledge base by allowing each incidence of unauthorized disclosure to inform us of emerging hazards. In this paper, we show how one can use the procedures for risk analysis in these industries to healthcare. Like these industries, we propose the creation of a centralized repository of possible adverse events.

Need for a National Incidence Database

We propose an objective methodology that utilizes cumulative experience, over time, thereby allowing the accuracy of the predictions to increase with use.  To accomplish this goal, we propose to build a blinded incidence database of unauthorized disclosures. Joint Commission on Accreditation of Health Care Organizations has created a similar database for sentinel events (e.g. medication errors or wrong side surgery).  If Joint Commission would consider privacy violations as a sentinel event, its database can serve as the repository for our proposed method.

Until a national repository is organized, we can rely on publicly available sources (complaints to Federal government, court cases and news reports) to organize one such database (See Figure 1).  The database is used to (1) define the focus of risk analysis, and (2) calculate the frequency of various risk factors among cases leading to unauthorized disclosures.    

Figure 1:  Sources of Data

If an ongoing national database existed, analysts can use it to expand their risk scenarios to include all known relevant risk factors. The moment a new hazard is identified in one part of the country; all subsequent risk analysis could include it -- providing a rapid method of improving risk assessments.  

Methods

The probability of unauthorized disclosure can be measured from a list of risk factors using the following formula:

p(U| R1, …, Rn )= ∑ i=1, .., n p(U | Ri) p(Ri)

Where:

n

is the number of hazards.

Ri

is the risk factor "i".

p(U| R1, …, Rn )

is the probability of unauthorized disclosures given various risks factors (vulnerabilities) in the organization.

p(U | Ri )

is the conditional probability of unauthorized disclosure given the presence of a risk factor in the organization.   This variable is calculated using the Bayes formula presented below.

p(Ri)

is the prevalence of the risk factor in the organization.   This variable is calculated from time-to-occurrence of the hazard within the health care organization (see below).

This formula is known as the law of total probability and it states that the probability of an unauthorized disclosure is the sum of all ways in which an unauthorized disclosure can happen from different risk factors within the organization.

We estimate the frequency of risk factors within an organization, p(Ri), by surveying key informants within the organization.  Since privacy risk factors can be rare, we assess the probability of their presence from the average time between reported occurrences of the risk factor:[1]

p(Ri) = 1 / (1+ ti)

Where,

ti

 is the average time between the reoccurrence of risk factor “i”.  These dates are available through survey of health care organizations.   

p(Ri)

 is the daily probability of risk factor "i".

Use of this formula assumes that the risk factor has a binomial distribution of occurring in which the probability of the risk factor is relatively rare but constant and independent from future occurrences. These assumptions may not be reasonable. For example, when organizations actively improve their security, then the assumption of constant probability is violated.  If the assumptions of binomial distribution are met or are acceptable as a first approximation, then time-between presence of risk factor has a geometric distribution. In a geometric distribution, the relationship between time-between events and probability of the event are given as per above formula.

Some risk factors are so rare that they may not occur during the observation period.  In these circumstances, then length of observation period can be used as a surrogate for time-between reoccurrences. This assumes that the risk factor would occur the day after the end of observation period and thus it provides an upper limit for the prevalence of the risk factor.

For an example of the use of the formula consider if we were to asses the prevalence of “physical theft of a computer.”  Suppose that our records show that such theft occurs once every three months, then the time between two thefts is 90 days and the probability of a theft for any day is calculated as:

p( Physical theft of a computer) = 1 /(1+91) = 0.01

This method of calculating prevalence of hazards stands in contrast to the arbitrary classification of risks by others. For example, the International Organization for Standardization (ISO) on December 2000 ratified the standard 17799 for management of information security.  The authors of this standard proposed to measure risk using the scale in Table 1. Table 1 also reports our approach to quantification of same scale. Clearly, the ISO 17799 standard does not accurately reflect the probability of the reported events. In fact, the correlation between ISO 17799 rating and calculated probabilities is 0.69, showing a moderate relationship between the two estimates.

ISO 17799 word assignment 

Frequency of event 

Rating by ISO 17799 

Calculated probability 

Negligible 

Unlikely to occur* 

0

0.0003

Very low 

2-3 times every 5 years 

1

0.0014

Low 

<= once per year 

2

0.0027

Medium 

<= once per 6 months 

3

0.0056

High 

<= once per month 

4

0.0333

Very high 

=> once per month** 

5

0.1429

Extreme 

=> one per day 

6

1

  Table 1:  Calculated Probabilities for ISO terms

* Assumes less than once per 10 years ** Assumes once per week 


We use the Bayes theorem to calculate the probability of unauthorized disclosure after the occurrence of a risk factor:

p(U | Ri) = p(Ri | U) p(U) / p(Ri)

Where:

p(Ri)

is the probability of observing risk “I”.  This is obtained from survey of health care organizations using time-to-occurrence of the risk factor.

p(U)

is the probability of unauthorized disclosure across institutions.  These data are calculated from National Incidence Database of Unauthorized Disclosures.

p(Hi | U)

shows the prevalence of risk factor “i” among unauthorized disclosures.  These data are available through National Incidence Database on Unauthorized Disclosures.

For example, if probability of unauthorized disclosure across organizations is 0.5% and the probability of observing computer theft is 1%, and the proportion of unauthorized disclosures attributed to computer theft is 2%, then the conditional probability of unauthorized disclosure following computer theft is 0.02*0.05/0.01 = 0.10 .

Results

In the absence of the National Incidence Database of Unauthorized Disclosures, we could rely on publicly reported privacy violations to show how our proposed method will work.  We identified publicly available reports of unauthorized disclosures from (1) review of complaints to Department of Health and Human Services regarding privacy issues, (2) Legal and news databases for reports of unauthorized disclosures. Table 2 shows the term used to search for unauthorized disclosures and the number of unique cases found:

Terms searched

Databases Searched

Records found

Number of unauthorized disclosures

Dates

Probability of unauthorized disclosure

Patient Confidentiality [Keyword] OR Confidential Medical Records [Keyword] OR Privacy [Keyword] medical records [additional terms] OR Privacy [Keyword] Medical Records [additional terms] unauthorized disclosure [focus]

LexisNexis Academic

47

2

01/01/03 -12/31/03

.005

Privacy of [Subject] Cases [Subdivision] OR Medical Records [Subject] Cases [Subdivision] OR Medical Records [Subject] Laws, Regulations and Rules [Subdivision] OR Hospital Information Systems [Subject] Safety and Security Measures [Subdivision]*

Health Reference Center-Academic Infotrac

141

8

01/01/90 -12/31/03

.022

US Dept of Health & Human Services HIPAA complaints

DHHS reports

22

16

01/01/03-12/31/03

.044

Direct reports

 

3

3

01/01/03-12/31/03

.008

Total: p(U)

 

213

29

01/01/90-12/31/03

.079

Table 2:  Frequency of Publicly Reported Incidences of Unauthorized Disclosures

*Also Includes: OR Business & Health, Feb 2001 v19 i2 p21 (Journal) OR Report on Patient Privacy, Oct 2003 v3 i10 p12 (Journal) OR Report on Patient Privacy, July 2003 v3 i7 p8 (Journal) OR Report on Patient Privacy, June 2003 v3 i6 p6 (Journal OR Report on Patient Privacy, Oct 2003 v3 i10 p12 (Journal) OR Computerworld, Dec 18, 200 p7 (Journal) OR InformationWeek, Dec 31, 2002 pNA (Journal) OR Modern Healthcare, Sept 15, 2003 v33 i37 p18 (Journal) OR Modern Physician, Nov 1, 2003 v7 i11 p2 (Journal) OR American Druggist, Jan 1999 v216 i1 p62(2) (Journal) OR AIDS Weekly, August 24, 1992 p16(2) (Journal)

For each case of unauthorized disclosure in our database, we described a risk factor that organizations could have reduced through security controls.  A comprehensive list of experienced hazards we identified through review of publicly available reports of unauthorized disclosure is available in Table 3. Obviously, security professionals may list many more vulnerabilities but these may not have yet occurred anywhere.

We estimated the probability of an unauthorized disclosure by examining the frequency of reports of unauthorized disclosures in legal, news and complaint databases. These frequencies respectively were: 0.009, 0.006 and 0.003 per day. The overall probability of unauthorized disclosure across all sources was 0.019.  Figure 2 shows the change in reported rates of unauthorized disclosures over the last decade.  There is a significant increase in the rates in 2003, when DHHS started collecting complaints


Figure 2:  Rate of Publicly Reported Unauthorized Disclosures Over Time

Our analysis of the reported privacy violations identified the risk factors in Table 3. What surprised us about this set of risk factors is its length. It is radically shorter than what a typical list of factors might look like.  Certainly other factors have been suggested but our review of observed unauthorized disclosures limited the list to items in Table 3. It is possible, and perhaps likely, that we are not aware of all cases in which unauthorized disclosures have occurred. Therefore our list of factors might be incomplete. It is also possible that other experts reviewing the same cases may arrive at slightly different set of risk factors. But the real point is that no matter who sets the risk factors and how many cases are reviewed, the number of risk factors will be relatively small because many risks can be imagined while few actually occur. Because relying on case histories reduces the number of risk factors, it radically reduces the time it takes to conduct risk analysis.

 

Using the risk factors in Table 3, we surveyed 3 organizations.  Table 3 shows the estimated frequency of risk factors in one of the surveyed organizations compared to remaining peer organizations.   Organizations can use the information to set priorities on which security problem they should address first.  Table 3 shows risk factors in which the organization is performing worse than the average of its peers.  This type of benchmarked information is important in helping organizations think through industry security and privacy standards.

 

Description of risk factor

Prevalence of risk factor in the organization

Prevalence of  security violation given the risk factor

 

Employee views paper documents or manipulates computer passwords to view records of patients not under his/her care

0.0003

1

 

Benefit Organizations or employers request employee information

0.0003

0.8805

 

Employees engaged in whistle blowing to uncover illegal or unacceptable business or clinical practices

0.0003

0.0201

 

Clinician using unsecured email environment

0.0003

0.1606

 

Employee removes patient records from secure location or workplace without authorization

0.0003

0.88

 

External infection of computers/password/network Systems (e.g. computer hacker)

0.0003

0.5888

 

Theft of computers or hard drives

0.0003

0.5867

 

Sale of patient records

0.0003

1

 

Blackmail/Extortion of organization or an employee

0.0003

1

 

Changes in custody or family relationships not revealed by the patient

0.0003

0.1472

 

Audit of business practices by outside firm without clinicians’ approval

0.0003

0.4416

 

Business Associate violates Chain of Trust Agreement

0.0003

1

 

Error in patient identity during data transfer to third party insurers

0.0014

0.0142

 

Caring for employees’ friends and family members and discussing the care outside of the work environment

0.0014

0.2202

 

Clinician gathers information from patients’ family and friends after the visit without the patient’s consent

0.0014

1

 

Patient using identity of another person to gain insurance benefits

0.0056

0.093

 

Patient records (paper documents) not kept in secure environment or sealed

0.0056

0.0592

 

Discussion of patient care with co-workers not engaged in care

0.0056

0.1218

 

Medical reports or records with wrong recipient information

0.1429

0.0405

 

Patient care discussed in a setting where others can easily hear

0.1429

0.0023

Table 3:  Predicting Probability of Violations from Prevalence of Vulnerabilities

 

Based on the risk factors present in the organization, an analyst can calculate the probability of unauthorized disclosure for the organization.  Table 3 shows both the prevalence of the risk factors within the organization and the relationship between the risk factors and unauthorized disclosure (estimated from NIDUD).  Health care organizations, regulators and health insurers can use this information to categorize the organization’s total risk category.

Discussion

This paper proposes a method for assessing risk of unauthorized disclosure within a health care organization.  It shows a mathematical formula for how to combine data on prevalence of hazards at health care enterprises with conditional probabilities estimated from the database of unauthorized disclosures. The advantages of the proposed approach are:

  1. It is not based on speculations regarding potential risks but on actual experienced incidences within the enterprise and across the industry.

  2. Probability of rare events are estimated from time-between-events which allows estimation of objective probabilities for events that occur once or twice.

  3. It provides a consistent quantitative estimate of risk. Our estimated probabilities may not be accurate to the last digit but they provide consistent processing of the information and thus the estimated risk can be compared to other institutions' risk factors and organizations can benchmark themselves against their peer.

  4. If repeated overtime, it measures the organizations' progress in reducing risk of privacy violations.

  5. It can be used to set premiums for HIPAA insurance, which can create a market incentive for more private healthcare transactions

For this approach to risk assessment to work, we need a centralized database of incidences of unauthorized disclosures. We did not have access to such a database but showed that the information can be assembled from public records of court proceedings, news reports, and complaints to Department of Health.  

Are the procedures described practical? The data needed at the enterprise level is minimal. Most organizations have a data set of security incidence categorized by type of hazard and catalogued by date of occurrence.  The creation of the cumulative national incidence database is more difficult. While we succeeded in creating it from public records, a more reasonable approach is to ask institutions to report these data to a central location, perhaps to the Joint Commission for Accreditation of Health Care Organization. The database calls for collaboration among health care institutions that may currently compete and perhaps may be concerned with the security of the data they share. We believe safeguards could be put in place that would protect individual institution’s identity. The benefit of such database outweighs its potential problems. In the absence of incidence database, risk analysis for HIPAA will not be consistent across organizations and may involve considerable effort and speculation about events that have not occurred or do not matter. Many organizations may not even conduct quantitative risk analysis and may settle for vague and arbitrary qualitative measure of low, medium and high risk. The incidence database can pool information from a large number of organizations so that it has sufficient cases to estimate various probabilities. The HIPAA legislation and final HIPAA rules already require organizations to maintain logs of every disclosure. The incidence database centralizes this information and allows the organization to use the information in a collaborative fashion to conduct its risk analysis. In the process of these collaborations, organizations learn about best practices.

Presentations

Questionnaire

A complete survey of privacy hazards can be downloaded by clicking here.

What Do You Know?

This section is under development. 

More

  1. Read about assessment of rare probabilities at http://gunston.gmu.edu/healthscience/730/ProbabilityRareEvent.asp
  2. The Geneva Papers on Risk and Insurance Theory (Kluwer)
  3. The Journal of Risk and Insurance Online
  4. Journal of Risk and Uncertainty (Kluwer)
  5. Risk (the official journal of the Risk Assessment & Policy Association)
  6. Risk Management Magazine
  7. An example of risk analysis using probability tree analysis. 

References


 

[1]           Centers for Medicare & Medicaid Services (CSM), HHS. Health insurance reform: security standards. Final rule. Fed Regist. 2003 Feb 20;68(34):8334-81.

[2]           Bedford T, Cooke R. Probabilitistic risk analysis: Foundations and methods. Cambridge University Press, Cambridge United Kingdom, 2001, 4-9.

[3]           Colglazier EW, Weatherwas RK. Failure estimates for the space shuttle. Abstracts of the Society for Risk Analysis Annual Meeting 1986 Boston MA, p 80, Nov 9-12, 1986

[4]           Keeney J et al. Report of the President’s Commission on the Accident at Three Mile Island, Washington DC 1979.

 


This page is part of the course on Decision Analysis, the lecture on Risk Analysis. This presentation was based on Alemi F, Shahriar B, Shriver J, Arya V. Probabilistic Risk Analysis (in review) . For assistance write to Farrokh Alemi, Ph.D. Most recent revision 03/28/2018. This work has been supported by grant from Critical Infrastructure Protection project. This paper reflects the opinions of the authors and not necessarily the policies and practices of their organizations.