Machine Learning (ML) will change three areas of the core processes we use in the life insurance industry today—paper and electronic application business, internal workflow, and underwriting medical data. The supply chain of life insurance distribution has for years promoted electronic forms for processing. While captive carrier models have the high adoption, 80 percent of distribution remains paper based. The good news is that even though paper is used to execute the agreement, if properly managed, an imaged document can be created and the original paper can be ultimately destroyed or stored. Internal workflow applications, commonly referred to as “Rules Based Workflow,” are programed applications managing routing, tasks, approvals, logic (if/then), limitations, people and exceptions, ensuring case management flows in an efficient manner. Medical input remains critical to underwriting so the best evaluation for an individual can be predicted to competitively price the policy. ML over the next few years will greatly impact these areas for the better.
Starting with Paper/E-App capture, Deep Learning (DL) and ML will be game changers, and this is why—color capture. In Part One of this article I discussed DL and it’s focus on processing color images, MNIST Training Set (TS) of grayscale images and DL models focused on biometrics. Today we challenge DL with the least optimum source image to process, the binary compressed TIF image; TIF images are limited to black or white pixels. Since the beginning of image capture our entire electronic documentation foundation is built on TIF images. The reasons for TIF remain as valid today as they did 30 years ago—quality reproduction output (paper-out), small byte count for bandwidth requirements and smallest storage footprint. So, we have a world of TIF images and a “slow to change” supply chain to overcome.
Paper/E-App capture will have to change from this day forward and change this inbound traffic to color. To take the greatest advantage of DL we need to start to capture with color images (i.e., PDF, JPEG, PRN, etc.). Delivering color images to your DL-CNN for text or handwriting recognition is your best chance of getting the best results. Getting producers, agents, registered reps, and advisors to scan in color, or mobile capture and save in color, to send in is all the change we need. This works because they are going to deliver those color images directly to your DL solution or DL service.
In the field of AI, people are working in a color world, no one is building Symbolic Learning (SL) or DL algorithms for black and white pixels; that would be like building PONG again. Therefore, if we expect great things from our technology, we must be willing to change. By introducing color into our documents, forms can use color to help page recognition, instructions for people and print or handwriting processing. Example: If my form when digitized (scanned) was the color blue and the person filled it out in black ink, when you do forms dropout you only remove the blue pixels; you preserve (black pixels) the original content which enables better recognition. Once the DL process is complete and you have received your data, convert the color image to a TIF image for storage. Conclusion: When building or selecting a vendor, ensure direct capture to the DL solution or DL service of color images.
Medical information and ML have several opportunities and challenges. The opportunity is that ML provides more perspectives to analyze data which will help when evaluating an individual. A true premise in Statistics and Probability mathematics is that you can predict the population but not the individual. Most systems are confined to three dimensions of perspective (i.e., lifestyle, health and family history). The one area you cannot predict is the ability of the individual to change. When negative lifestyle and health behavior changes to the positive, this individual changes to a new carrier/product for better coverage value, while the opposite change goes unnoticed. ML now offers the opportunity to add on new data sources available and predict the ability of the individual to change. New sources like social media, electronic health records, DNA screening, internet activity, buying trends, twitter content, criminal violations, fraud detection and more will help predict the individual aptitude or tendencies to change.
ML models are pliable. One model can be used for many processes and even across different industries. The point is, one ML model could automate an entire company. UCLA developed a ML algorithm to predict earthquakes around the globe. The same ML algorithm is used by the Los Angeles Police Department to predict crime. In both cases, they used historical data to train their models and have produced significant results. LAPD has experienced a reduction of 33 percent in burglaries, 21 percent in violent crime, and 12 percent in property crime across the area where the algorithms are being applied. LAPD built a TS incorporating data going back as far as 30 years. The next effort underway is to process more sources to fine tune the ML parameters.
This ties back to change in medical information because finding the right model is key and next is your TS. This now becomes a challenge because the Electronic Medical Record (EMR) is 90 percent image based. The good news is that the Electronic Health Record (EHR) is digital. The bad news for EHR is that there are no standards; interoperability between vendors is nonexistent. EMR requires a process of transcribing images to usable data to build your TS. All the Attending Physician Statements (APS) and Part 2 medical forms need to be converted into data. These are one-time projects called “Dark Data” and the trend in the life insurance industry is converting Part 2 forms.
Carrier Part 2 medical forms is another area of change that needs to happen to make ML/DL effective. Carriers and others paying for medical exams need to stress to the examination companies to “Follow the Forms’ Instructions!” Working with many Part 2 forms we see shortcuts which to humans may be acceptable but to ML are accuracy killers. Example: Part 2 checkbox groups where the author strikes a diagonal line through the group intending No as the answer for all the checkboxes. This line now passes through some checkboxes making them true to DL when the intention was for false. Without a level of quality control of the form, this individual could lose their chance or at least add additional expense to the carrier’s process. Part 2 form designers need to add more space when collecting explanations to requested questions. Designers should provide data capture examples (i.e., MM/DD/YYYY, $1,000, Group checkbox selection, etc.) and allow space to write equivalent to 30 font size.
Digitizing the APS has challenges, but with ML we see the opportunity to transcribe for about the same cost of summarization. When underwriting ML matures, the APS data will best reveal the individual. Predictive Analytics will become a competitive edge if focused on the individual and not the population. In the medical industry their ML efforts are focused on quality of care. The most used vendor service for doctors and medical providers is “UpToDate” a Wolters Kluwer company keeping professions up to date with the latest clinical support resource with improved outcomes. Medical professionals also want “Natural Language Processing” (NLP) because they know there is so much more data collected if the doctor—patient encounters could be recorded and processed. Unfortunately, there is very little progress with NLP and medicine, but this is the hottest AI area according to Robert Wacther, author of The Digital Doctor and others. To life underwriters I would recommend working with their summarization partners and get the ML projects going to build TS, something small but that still has a positive impact, and start cheering for medical NLP.
Rules Based Workflow Is Dead—Long Live Machine Learning! Over my career I have never seen a Rules Based Workflow (RBW) produce a ROI. I have been involved in many RBW projects that took many hours of consulting, configuration, programming and training wrapped in much debate on how the flow really works. A year or two in development, finally launched, the standard comment was “Great—we’re only two years out of date.” RBW is like the newspaper business—their product has a 24-hour lifespan. Change it and the staff required to manage it simply stops, users pushout work-arounds, and then the entire RBW is killed in the first reorganization chance that comes along.
ML is a different paradigm. You don’t configure or program rules. ML writes the rules on the fly based on its data inputs. ML will learn from all three training events (Supervised, Unsupervised, and Reinforcement) and, based on input, change the workflow. Let’s walk through such an event like routing information for approval. Jillian works within a New Business processing shop; she and her boss Lara are the only ones who are authorized to approve a type of application. Jillian takes a one-week vacation and therefore won’t be available to approve, leaving Lara to watch her queue or get sysops involved to change that route for a week and then change back. ML would learn from email that Jillian is out of office. ML knows from the existing TS that this type of application is routed to Jillian 90 percent of the time and Jillian approves the application compared to Lara’s 10 percent of the time. ML could have an input from the email solution for Jillian’s “Out of Office” notices and change the routing to Lara’s queue and back to Jillian’s when the input changes. The point is, new workflow solutions will be manual with ML taking over as the data and TS mature. This is where the data science comes in, reducing program development for data development and, in time, the workflow will maintain itself.
I talked about three areas AI and ML can impact the way life insurance is processed today. Image capture, future back office systems and medical underwriting. Of those three, there is no question the industry is focused on medical underwriting, the real opportunity for competitive products. Next, ML workflow for its inherent focus on your processing and your product optimizing resources. Leave “image capture to data” to vendors—it offers little, if any, ROI when built internally. Final point: Focus your ML efforts on your largest and most strategic return, and outsource the rest.