The AI Age is nigh upon us. Repent, for the end is near? Hardly. But, as with the Stone Age, the Bronze Age, and the Industrial Age, those who fail to adapt do so at their own risk. This inevitability extends to medicine and physicians in the face of AI. But this editorial is not about the pros and cons of AI itself. Rather, it concerns an insidious present-day practice by some physicians in which responsibilities are relinquished to a process that prioritizes an unseemly profit motive. It may be a glimpse into the future where even greater temptations may be offered to physicians by widespread AI.
In my own specialty, I have become increasingly disheartened by many physicians' reliance on sleep study interpretations generated by automated programs, to the exclusion of any personal oversight. Even worse, a growing reliance on scoring from ancillary healthcare members has, in my opinion, become disturbingly widespread. I can only comment on my area of expertise, though this explanation may apply to other, mainly diagnostic specialties, rather than to medical specialties that rely on procedures and interventions.
This goes beyond simply adapting to an “information age,” in which reliance on technology and scientific achievements advances patient care. When reliance becomes absolute dependence on and exploitation of interpretation by skilled but lesser-trained colleagues, it becomes an abdication of physician responsibility and an unacceptable standard of care. Too often, I have seen physicians simply sign off on interpretations without a glance at the actual data, accumulating remarkable productivity in a remarkably short period of time.
If referring physicians and patients still obtain direction and information while efficiency is maximized, isn’t this simply a win-win for all involved? Not when meaningful interpretation of any test involves subtleties and nuances outside the capabilities of algorithms or abbreviated education. Whether gleaning information from a polysomnogram, an unattended home study, a simple oximetry study, or a chest X-ray, the absence of certain expected findings can be as informative as their presence. Such insight can be unappreciated by a registered polysomnographic technologist or an automated algorithm.
Further subtleties can be overlooked if a sleep study is not interpreted in the context of the patient's information or any unexpected findings. A patient's complaint of abnormal behaviors during sleep may require a detailed electrophysiologic evaluation for an underlying seizure disorder or a disorder of arousal (e.g., somnambulism). Such a deep dive would likely be beyond the scope of a respiratory therapist or an AI algorithm.
If an “unrevealing” sleep study undertaken to evaluate specifically for sleep apnea is taken at face value, subtle EEG or electromyography findings suggesting fibromyalgia or a disorder of hypersomnia may be missed or simply ignored. Periodic limb movements, arousals with heart rhythm irregularity, or repeated nocturia may be regarded as incidental polysomnographic clutter rather than felt to be informative to a referring physician coping with a patient’s insomnia.
Possibly most important, I have seen very mild sleep apnea, signed off without actual interpreting physician oversight, generate a recommendation for specific therapy such as CPAP. Such a recommendation is generated by an algorithm and may expose the patient to unnecessary titration or even long-term, unnecessary, and frustrating CPAP treatment. Not all levels of sleep apnea require treatment, but this caveat may not be recognized by an AI algorithm or a respiratory therapist responsible for unmonitored interpretation.
Diagnosis does not necessarily equal treatment, but too many interpretations suggest otherwise when completed without educated or trained insight. Poor sleep and nocturnal frustration are shameful outcomes of a process designed to improve sleep, yet they occur all too frequently.
At the risk of a more intangible moralistic objection, reaping income from a simple, unsupported signature lessens the inherent worth of a patient and violates the sacred responsibility of managing that most valuable of concerns — the health of a patient and his or her quality of life. Were any such physician on the other side of the white coat, they would expect nothing less than undivided attention from those providing care. Personal curbside peer review, or even a formal second-opinion review, would likely be sought. Why should the everyday patient without such means and influence expect and receive less? Prioritizing profit over hard data has become uncomfortably acceptable.
“Uncomfortable” observations shouldn’t equal “untrustworthy” observations; paper trails exist whether actual or digital. Criticism shouldn’t reflexively be regarded as a self-righteous accusation, nor should it be resented as an indictment against the majority of dedicated peers. Taking the path of least resistance and cutting corners, especially when there is financial motivation, may go unrecognized by the lay public.
But at least with regard to my particular specialty, it is easily identified by those on the inside, cheapening the profession and reducing trust in supervising practitioners. Furthermore, affixing a signature to data implies responsible oversight and leaves the signatory legally liable for errors or overlooked findings.
Appropriating an algorithm's or lesser-trained (and lesser-reimbursed) clincian's interpretation as one’s own may be a precursor of medicine under the even more tempting efficiency of AI. Virtually all professionals, including physicians, are vulnerable to time- and effort-saving temptations, especially when they lead to financial gain. At what point does responsible augmentation of effort and expertise become exploitation? Perhaps when “augmentation” becomes wholesale substitution of effort and oversight? Or when ancillary healthcare workers put forth invisible work and effort without appropriate compensation, which is then billed for by a superior? Like the observations already put forth in this article? At least AI performs without such reward disparity or exploitation (at least until it becomes self-aware and unleashes Skynet).
The solution seems obvious but paradoxically resistant to action … i.e., blowin’ in the wind. Random or mandatory peer-on-peer institutional self-policing would likely only provoke resentment rather than compliance. Financial repercussions for subpar performance would likely reduce exploitative activity but would rely on random peer or AI review with the same attendant adversarial oversight. Personal self-policing is called for, although such a plea is all too commonly sought and ignored.
The current practice of medicine is most assuredly burdened with time constraints, clinical and administrative stresses, patient scheduling demands, and reimbursement shortcomings. Our profession is thus fertile ground for relinquishing the insight and expertise from years of education and experience to less discerning and less compensated staff or rigid scoring algorithms. It may already be out of our control. Programs are in place to record time spent on a specific endeavor, clearly differentiating time spent solely affixing a signature from time spent responsibly evaluating data or test results. Such surveillance will only become more intrusive and specific.
Voluntary change must come from within, before even more invasive, widespread involuntary change is mandated from without.
What are some of the ways you see automation being helpful or harmful in healthcare? Share in the comments.
