This year, HHS has put out two major and broad rules that impact innovative medical technology. In both cases, the published cost-benefit analysis was a mess. The two rules are an FDA proposal to regulate lab developed tests (proposed 9/2023), and an HHS ONC rule on health system information systems and AI (proposed 4/2023, final 12/2023). Both rules divided stakeholders. The lack of an ability to propose credible cost-benefit data in favor of either rule, which is the duty of proponents and advocates, is noteworthy.
FDA - Terrible Cost-Benefit Discussion
Most familiar to readers here is the FDA LDT rule, in proposal, commens closed December 4. Among other problems, the rule was extremely vague in some key aspects, and had a god-awful cost benefit analysis. I wrote about that early in the fall in a white paper. Among many commenters, ACLA emphasized the awful cost-benefit analysis with a 25-page economist's appendix to their comment letter. (See an AI tutorial on Prof. Carrigan's report, here.)
HHS ONC releases EHR-AI Rule ("HTI-1")
I regret I didn't stop and pay enough attention to an HHS (ONC) proposal to upgrade their regulation of EHR's and AI, issued in proposal form last April. See an outstanding white paper on the ONC proposal issued in June by McDermott Will & Emery, here. Turning to the new final rule, see Fierce Healthcare here, ONC home page for the rule HTI-1 here.
Others were paying attention - see descriptions of major comment letters from the likes of Amazon, Google (here at Becker's Health It, here at Stat). Each was just 5pp. (For an AI analysis of the 2 letters, here.) (UHG had a vehement, longer letter here.)
The final rule appeared in Fed Reg on 1/9/2024 - 89FR1192 (247pp). Here.
HHS ONC Cost-Benefit Discussion- A Tissue of Confusions
Having been appalled from A to Z by the cost-benefit analysis in the FDA LDT proposal, I looked at the cost-benefit analysis in the HHS ONC proposal.
The ONC proposal matches the final, because, in contrast to the barrage of criticism on this topic at FDA-LDT comment period, there was NO COMMENT to most of the equally awful cost-benefit analysis served up by the writers at HHS ONC. My head was spinning.
They make predictions of benefits for sepsis management of $432M (over ten years when there are $200B of sepsis admissions) and $600M of saved general admissions (in a ten year period with a base cost of $240B of a type of admission) (proposal table 19). (See discussion at 88 FR 23889 ff). Software costs range up to $335M (proposal, table 18). When they get to summary figures (proposal 88 FR 23903) it's $742M for total annual IT costs and net annual benefits the suspiciously round number, "1 billion," allowing the whole thing to pass the test for "unfunded mandates" (88 FR 23905).
Of course, if hospitals (with passed-on IT developer costs) incur $0.7B in expenses and have $1B less in DRG revenue, they are $1.7B down (in the red), not $300M ahead (in the black).
What Staff?
ONC also makes unusual assumptions about the labor require, as either management analysts (about $45 per hour before overhead) and software analysts (about $55 per hour). But the rules involve new and extremely complex conceptual medical risk-benefit reports and other reports on every piece of software (see McDermott white paper). The work of doing this is unlikely to be within the grasp of an entry-level programmer or analyst. (Plus, AI software experts are hardly $55 an hour!)
Overall, I was surprised that the important cost-benefit section of the ONC EHR AI rule attracted few to no comments. If you don't know the benefits exceed the costs...?
Another Kinship: Statutory Overlap is Underdeveloped: FDA-CLIA vs ONC-FDA
Another topic in itself. The FDA LDT proposed rule collides with CLIA in many ways that the FDA left wholly for future "filling in the details."
Same deal here, but in a final rule. The ONC EHR AI rule, appears to have many points of collision with FDA regulation of medical software. Writers complain about this (Final Rule, Inspection Copy, 200ff). However, ONC handles these concerns and some vague inline sentences and a couple uncommented crosslinks to FDA documents.
Updates Sexual Orientation and Gender Identity
The final rules update on sexual orientation and gender identify in several contexts (45 CFR 170.207).
(o) Sexual orientation and gender information—
(1) Standard. Sexual orientation must be coded in accordance with, at a minimum, the version of SNOMED–CT® codes specified in paragraph (a)(4) of this section for paragraphs (o)(1)(i) through (iii) of this section and HL7 Version 3 Standard, Value Sets for Administrative Gender and NullFlavor (incorporated by reference, see § 170.299), up until the adoption of this standard expires on January 1, 2026, for paragraphs (o)(1)(iv) through (vi) of this section, attributed as follows:
(i) Lesbian, gay or homosexual. 38628009
(ii) Straight or heterosexual. 20430005
(iii) Bisexual. 42035005
(iv) Something else, please describe. nullFlavor OTH
(v) Don’t know. nullFlavor UNK
(vi) Choose not to disclose. nullFlavor ASKU
(2) Standard. Gender identity must be coded in accordance with, at a minimum, the version of SNOMED–CT® codes specified in paragraph (a)(4) of this section for paragraphs (o)(2)(i) through (v) of this section and HL7 Version 3 Standard, Value Sets for AdministrativeGender and NullFlavor (incorporated by reference in § 170.299), up until the adoption of this standard expires January 1, 2026, for paragraphs (o)(2)(vi) and (vii) of this section, attributed as follows:
(i) Male. 446151000124109
(ii) Female. 446141000124107
(iii) Female-to-Male (FTM)/Transgender Male/Trans Man.
(iv) Male-to-Female (MTF)/ Transgender Female/Trans Woman. 407376001
(v) Genderqueer, neither exclusively male nor female. 446131000124102
(vi) Additional gender category or other, please specify. nullFlavor OTH
(vii) Choose not to disclose. nullFlavor ASKU
(3) Standard. Sexual Orientation and Gender Identity must be coded in accordance with, at a minimum, the version of SNOMED CT® codes specified in § 170.207(a)(1).
(4) Standard. Pronouns must be coded in accordance with, at a minimum, the version of LOINC codes specified in 170.207(c)(1).
##
I posted in a sidebar, an initial back-and-forth AI discussion of the ONC EHR AI rule - here.
28 companies signed a general commitment to develop AI responsibley; Reuters Dec 14 here.
