Mr. Chairman and Members of the Subcommittee:

Thank you for providing AMVETS (American Veterans) the opportunity to submit our views regarding the use of artificial intelligence in VA claims processing.

The claims backlog that plagues the Veterans Benefits Administration (VBA) has been a great concern for veterans, and AMVETS is pleased to see the Committee on Veterans Affairs is taking the time to genuinely study this issue so long-lasting, effective changes can take place. AMVETS believes the use of artificial intelligence (AI) could greatly reduce the time involved in processing disability claims. The use of AI to reduce the amount of data provided to reflect only the information that is necessary to make timely and accurate decisions is not a new idea. The government has been using AI to select applicants for federal positions and AI has being used in occupational health services to determine compensation for private and government facilities. So the question is not can AI be used to assist VA in the claims process, it is, to what extent can and should it be used?

AMVETS generally supports the idea of using AI to improve the VA’s claims processing system with some reservations. Integrating AI into the VBA to assist in the disability claims system could greatly improve the efficiency of the claims process or it could exacerbate the problems in the current method of adjudicating a claim. The use of AI would bring about a 100% electronic method of record keeping, which would provide easy access to records at all phases of the claims process. This would prevent the loss or misplacement of information.

AI has the potential to significantly increase speed and accuracy during the triage and pre-determination stages of the claims process. At these points of the process, over and under development of claims happens all too often. With AI reviewing all of the incoming medical records, compensation and pension (C&P) exams that need to be ordered would be identified. This could prevent redundancy in examinations and ensure all necessary exams are requested and conducted, preventing the Rating Veteran Service Representative (RVSR) from beginning their work only to realize that additional C&P exams will be required. The rater now has the choice of either doing a partial rating, or stopping the rating, requesting the appropriate C&P exams, and completing the rating decision when the C&P exams return.  Normally, this is a sixty day process.

Disparity in the rating system from one regional office to another has become a major issue in disability claims compensation. The use of IA could, if developed properly, provide the rater with the appropriate diagnostic code for the determination of the percentage of disability for which a claimant qualifies. This is because AI also has the ability to recall all diseases and medical conditions and almost instantaneously compare all of the veterans' conditions against its database. Because of this, rare diseases and conditions that rating specialists see infrequently, such as keratoconus, will not be overlooked. Conversely, AMVETS has a concern that too much reliance on AI could result in not allowing logic to be used in determining the qualification of disability, something only a human can provide. So a line would need to be draw on how much responsibility will be given to the AI.

However, the level of complexity of the AI would require brings about concerns on its own accuracy. The AI software would either have to be able to read hand-written documentation or physicians would have to type all of their findings. Also, there are certain tests that are represented in diagram or table form, such as the loss of field of vision test and the results of audiologist exams. The AI would need to understand and interpret these types of exams or be able to recognize their presence and prompt the rating officer to review them. This issue begs the question, will AI be grandfathered or will it have a starting point? If there is a grandfather provision, the AI will need to recognize medical terminology that has a tendency to change over time.

To utilize artificial intelligence properly, the terminology inputted into the system would have to match the terminology the system recognizes. This brings to light the inconsistencies between the checklists contained in the Disability Evaluations Examination Worksheets used by VA physicians to conduct compensation and pension (C&P) physical examinations and the criteria contained in 38 C.F.R., Part 4, SCHEDULE FOR RATING DISABILITIES.  In many cases, the descriptive words do not match. AMVETS’ concerns are that AI software is not smart enough to provide accurate results when a physician’s report does not exactly match the description of the rating even though the intent of the report describes the disability and the rating veteran service officer (RVSR) could interpret the meaning of the report though logic.

There is a fundamental disconnect between the check list that prompts physicians on the Disability Evaluation Examination Worksheet that is used during the Compensation and Pension Exam, and the description in 38 CFR Part 4 that is used by the RVSR to determine the percentage, if any, that will be granted to a claimant. For AI to be an asset to the claimant, the physician’s write-up must match the language used in 38 CFR Part 4. Under the current system of evaluation, a rating officer can determine the intent of a physician’s evaluation.

To receive a 30% disability rating for PTSD, 38 CFR Part 4 states:

Occupational and social impairment with occasional decrease in work efficiency and intermittent periods of inability to perform occupational tasks (although generally functioning satisfactorily, with routine behavior, self-care, and conversation normal), due to such symptoms as: depressed mood, anxiety, suspiciousness, panic attacks (weekly or less often), chronic sleep impairment, mild memory loss (such as forgetting names, directions, recent events).

A claimant may have these conditions and they may be documented in the physician’s evaluation. However, if the physician states the patient has some memory loss, rather than listing it as “mild” memory loss, then AI may not discern between the words or phrasing used by the physician and what it is looking for as it relates to 38 CFR Part 4. These logical word choices could easily result in the claimant receiving a lower or no rating for a disability. For artificial intelligence to be applicable it would be necessary to develop a Disability Evaluation Examination Worksheet that is based on the language used in 38 CFR Part 4 so semantics or synonyms would not reduce or prevent a claimant from receiving disability compensation. This would be easiest to facilitate on ratings of lost limbs or joint problems and become more complex when dealing with issues that have multiple variables such as PTSD or TBI.

Again, AI could greatly benefit VA and the claimants if the system was allowed to work in all phases of the claims process and the information received by the system was completely and properly reviewed. Also, it is important to maintain the ability to intervene with human logic when necessary.   

Mr. Chairman, this concludes my testimony.