Automated Analysis of Competing Hypotheses for Diagnostic Medical Decision Support

It is with great pleasure I publish a project, hopefully to the benefit of society.  With my co-author Kindall Deitman, I have published a Diagnostic Medical Decision Support System which implements analysis of competing hypotheses.  It does so in the form of two new machine learning algorithms, the Bassett Deitmen Training Algorithm and the Bassett Deitmen Query Algorithm.  For those who follow this blog, it should come as no surprise that the underlying model is a directed graph.  For more information on the model, please reference the paper: Graph-based Diagnostic Medical Decision Support System.  You may watch our presentation of the project, where we answer a few questions as well, here.  If you would like to test or contribute to the model, it is licensed for non-commercial use here.  Should the need arise for a commercial license, please contact myself or Ms. Deitmen.

It is truly our hope that this model can benefit society as a whole.  Medical diagnosis is something that happens, regardless of the skill of the practitioner.  Whether it is a doctor with decade of experience, a new physician's assistant, a nurse in a small town or village, or even a bystander with little more than first aid training, when medical diagnosis is needed, it does not wait.  This project hopefully puts the experience of innumerable medical practitioners at the fingertips of those who need it most.  Additional, it hopefully brings a means of 'jogging the memory' of experienced medical practitioners who may have trouble remembering the obscure illnesses they rarely see.

There is certainly room for improvement.  A rather long list of updates and improvements already exists as we look to increase the utility of the project.  Nor is the approach constrained to medical diagnosis.  The training and query algorithms can support any situation in which observations or signals must be used to prioritize hypotheses as to the cause.  Still, there is a real and present need for such tools to help medical practitioners.  In an age where EMRs hold an incredible wealth of knowledge, it is our hope that this project may allow us to truly begin to harness it for the good of all.

0’day Campaigns for Everyone!

Hop over to the Verizon security blog to read my most recent O'day Campaigns for Everyone! (or why every attack now a'days looks targeted).  It's amazing what you can do with the DBIR data!

Association Rules

Check out my blog on Association Rules over at the Verizon Security Blog.  It's definitely an under-utilized area of machine learning.

Standardized Data Trees (UN M.49, ISO 3166-1, ISO 3366-2, country population/area, NAICS)

To help in aggregating, comparing, and validating data, I've created a pair of data of graphs which represent tree hierarchies of standard formatted data.

The World Graph contains:

• UN M.49 country codes
• ISO 3166-1 alpha-2 and alpha-3 codes from pycountry
• ISO 3166-2 country subdivision codes from pycountry
• World Bank country population data
• World Bank country geographical area data
• Aggregation of population and geographical area data at higher levels of the graph
• (Some population and geographical data filled in from wikipedia where missing from the World Bank)

World Graph Visualized

The NAICS Graph contains:

• 2012 NAICS codes in hierarchical form
• Percentage of the graph under each node

The percentage under the graph in the NAICS graph and the aggregate population/geographic area allow two things:

1. Provide an amount as a dimension for data coded in these systems, (whether that amount be percentage of NAICS codes, population, or geographic area).
2. Provide a means of comparing the similarity of two records by finding the Lowest Common Ancestor (LCA) and retrieving the score from that node.  The greater the score, the greater the distance between the nodes.

The hierarchies can be used for validating data as well as comparing things (as above) that are not on the same level of the graph.  For example, a US state could be compared to the country to South-Eastern Asia.

None of this is groundbreaking, but hopefully some find the graphs of use.

The Opportunity to Create

We work in a great profession in information security.  Unlike other professions which are bound by the physical world, we work with near limitless scope as infosec's context is not just spread out over the physical world, but also the digital.  In addition, we work in a profession whose challenges are not just static, caused by physical/technical constraints, but also dynamic, caused by the competing interest of different people.  However, that also hinders us in succeeding in our profession.

Information security has always had a combative context.  It's understandable as there is a clear offensive side, a clear defensive side, and rational actors existing on both sides.  We think about solutions in the context of winning the conflict.  This leads us to look for solutions based in force.  Either the force to overcome the other side's defenses or the force to absorb the other side's attacks.

There is another way though.  Instead of thinking of conflict, we can think of building something that simply transcends the conflict.  The same way a dancer compliments their partner's movements rather than forcing their partner to do what they want, we can think of information security as the opportunity to create something that transcends the combat; to create something that makes the combat a suboptimal solution to the goals of those participating in it.

I can't say I know what those solutions are.  I'm sure they are much harder to find than simple us versus them solutions.  however I think the transcendentals are a good starting place:

1. Goodness: Is the solution good, (and not just in the moral relativistic sense that it is good in my own context, but in all stakeholder contexts.)
2. Truth: Is the solution true, (again not just in the relativistic sense.  It must be true for all stakeholders.)
3. Beauty: Is the solution beautiful to all stakeholders.

However, the transcendentals are very abstract concepts to apply.  In our day to day work, the following tenants may be much easier to test:

1. If we are thinking of how a solution helps us gain an advantage over someone or beat someone, it is not the right train of thought.
2. Finding solutions should include thinking about all stakeholders on all sides of the conflict and their needs.
3. Finding solutions should include thinking about how the capabilities of all stakeholders can be integrated to create something greater than the sum of the parts.
4. The solutions may not be technical in nature and may require the inclusion of stakeholders with non-technical skills to implement.
5. We should be prepared to compromise and sacrifice to find the solution.

Hopefully by considering these tenants as we think of how to solve information security problems, we can find solutions which transcend the daily conflict of information security.  Hopefully we can find solutions which prevent conflict not because of the risk of losing is too great, but because there is no incentive to engage in it.

So the next time you are trying to solve an information security problem, test your approach to finding a solution against the tenants above.  If you find that your approach is inconsistent with the tenants, consider what you could do to meet these tenants.  The better a solution meets these tenants, the more likely it is to be a long lasting solution.

I Made a Million Models

Check out my blog over at Verizon Security:  I Made a Million Models.

When is an Intelligence Feed Record New?

Head on over to the Verizon Security blog to check out my new post on when an intelligence feed record is new.