I’m a bit of a trivia nerd. In fact, I play trivia with a group of friends every week. We do alright, and obviously there are good weeks (I mean, we keep going back) and then there are bad weeks. I play team captain for our group. The responsibilities of team captain are to record our progress (each question gets a wager based on our confidence in our answer) and recording the success or failure of each question in a running total, and helping to marshal the resources of the team (points, knowledge bases of the players, ranking answer likelihood, etc.). The final trivia question of the night is a challenge. Each team is given the question that requires four answers in rank order, usually. When turning in one’s response, a point value between 1 and 15 is assigned to the answer. If any part of the question is wrong, the wager is subtracted from the team’s total score. If the answer is 100% correct, the team gains the wagered points. So it’s no surprise that I would be really intrigued by Watson, a supercomputer that was able to best two of Jeopardy’s greatest champions in a tournament back in 2011. Research into Watson is really interesting.
Watson was trained to respond only when a certain threshold had been met in the likelihood that Watson was correct in its assumptions. This confidence was determined based on cross-referencing the available answers and identifying the highest scored answer based on a number of algorithms. While Watson is not right 100% of the time, its significant domination of the final score ($77,147 vs. 2nd place’s $24,000) is no small feat for a computer responding to natural language, searching natural language information, and culling a response to an “open-domain” question.
Game State Evaluation
Part of the programming behind Watson required not just an understanding of the likelihood that Watson was right, but also what the potential for gain or loss in relation to the other players might be. Because Jeopardy includes wagering for daily doubles and Final Jeopardy, Watson had to strategically wager in relation to the likelihood not only that it was right, but also what it would mean if the other players were right or wrong. This is well-illustrated by the final wager that Watson placed in response to the final jeopardy question, which was $17,973. This is a statistically-determined wager based on total game state evaluation at the time of this final question using the above variables.
While there is plenty of room for argument as to whether or not Watson is thinking, there is absolutely no question as to whether Watson is logical. As I’ve mentioned before in a couple of articles related to the work of Daniel Kahneman, (if you haven’t, make sure to check out Thinking Fast and Slow) human rationality is very rarely very rational. This is due to a number of intervening variables that interrupt our ability to make rational decisions all of the time. These “biases” can be intentionally or unintentionally applied during the decision-making process. While Watson has a number of heuristics, no-doubt, built into its logical processing, it is probably not as likely to respond to cognitive biases such as anchoring, duration neglect, and certainly curse-of-knowledge as seen in its commanding performance in the Jeopardy games.
Decision Support Systems
Watson is now being used in a number of healthcare applications assisting in the support of clinicians as diagnostic support. Watson is not making decisions, but it is able to cull the plethora of information available in the medical field to provide confidence-rated responses to data that is provided regarding a patient. This marks a big step for the advancement of Health IT as we can standardize clinical response to symptoms, and stabilize health information as it is consolidated into big data stores. And because Watson is able to learn as it answers and receives feedback as to success and failure based on those responses, Watson can only get better at diagnostic prediction and likelihood of treatment success or adherence based on the results of those treatments.
What does this mean for making people healthier?
There are some obvious benefits to these advances in Health IT, but one of the things that may not be fully clear yet is the application of Watson to understanding more about human behavior. While Watson can absolutely tell a clinician the likelihood of a set of symptoms’ association with a given disease, I’ll bet Watson can’t tell you how the patients’ family impacts their overall wellbeing through behavior reinforcement. If Watson knew who the patients’ workout buddy was, Watson might be able to help identify with a high confidence whether that workout buddy was a statistically-sound partner in the overall health management of the patient. Further, Watson would be able to weigh in on the evaluation of treatment adherence based on real-time data pouring into the health record for the given individual. This is the game state evaluation of the health of the individual in a real and meaningful way. With this, a total and complete understanding of the long-term treatment of chronic conditions (and even more important to the salutogenic framework that I’ve discussed previously in this blog series, total health production) through the understanding of actual human behavior devoid of the clinical separation from reality is the “social human” version of epigenetics that will become more useful in the coming years. This is where the data comes to life.
To our health,
To stay ahead on topics related to this, follow me on Twitter @dz45tr