Mid-week roundup: Promising Langone AI/LLM predicts hospital readmits; Huma gains FDA 510(k) Class II clearance; telepsychiatry’s challenges; layoffs/asset buys/losses from 23andMe, Cityblock, Thirty Madison, Butterfly

June 14, 2023 | by

New York University’s Langone Health’s large language model (LLM) accurately predicting hospital readmissions, more. NYU’s academic medical center NYU Langone Health has developed an LLM using medical language, NYUTron, from unstructured clinical notes in patient records, then fine-tuned it across a wide range of clinical and operational predictive tasks. The dataset was immense:  ‘NYU Notes’ covers 7.25 million clinical notes (for example, radiographic reads, history, and physicals) from 387,144 patients across four hospitals, and more. According to their study published in Nature on 7 June, it was tested for predictive ability in five areas: 30-day all-cause readmissions, in-hospital mortality, comorbidity index, length of stay, and insurance denial. The NYUTron system in testing has achieved results improved over conventional structured models’ baselines. From the Nature study:

  • For 30-day readmission prediction, it had a median area under the curve (AUC) of 79.9% ± 0.168% with a 5.36% improvement
  • On in-hospital mortality prediction, NYUTron had a median AUC of 94.9% ± 0.168% with a 7.43% improvement.
  • On comorbidity index imputation, NYUTron had an OVR median AUC of 89.4% ± 0.275%
  • On binned LOS prediction, NYUTron had a median AUC of 78.7% ± 0.179% with a 12.3% improvement 
  • On insurance denial prediction, NYUTron had a median AUC of 87.2% ± 0.246% with a 14.7% improvement.

In a test of the system during January-April 2022, the system analyzed 29,286 discharged encounters, with 3,271 patients (11.17%) returning within 30 days. NYUTron predicted 2,692 of the 3,271 readmissions (82.30% recall) with 20.58% precision. Also HealthcareITNews

London-based Huma (the former Medopad) gained US FDA 510(k) Class II clearance for their Software as a Medical Device (SaMD) platform. This is defined as disease and age-agnostic digital health pathways through which data are collected from patients for self-management or to be assessed remotely by healthcare professionals. Huma also recently obtained EU MDR Class IIb approval and with Health Canada through the FDA’s joint eStar program. Huma’s tech also includes remote patient monitoring (RPM) systems and companion apps to enable disease management, with third-party device integration. For providers, the platform hosts AI algorithms that use automated data analytics to support screening, diagnosis, dosing recommendations, clinical decision making, and prognostication for identification of at-risk patients and early intervention. In 2020, Huma acquired BioBeats and TLT; more recently, last year iPLATO patient engagement and in January clinical trials data specialist Alcedis.   Huma release, Mobihealthnews

The growth of behavioral health has come to a screeching halt with the demonstrated abuse of online prescribing, then the US Drug Enforcement Administration (DEA)’s uncertainty around controlled substance prescribing. This interview with the CEO of Array Behavioral Care, one of the Ur-companies in telepsychiatry (1999, originally InSight Telepsychiatry and Regroup Telehealth), points out how the DEA’s post-Public Health Emergency (PHE) policies around Schedule II and higher teleprescribing disrupted their operations. The flexibilities established during the PHE have been waivered to 11 November, though a final rule must replace the temporary extension rule and comply with the Federal Ryan-Haight Act [TTA 11 May]. Other issues addressed are dealing with medical affairs (clinical licensure, primary source credentialing, facility privileging, and payer enrollment), and the potential for AI to create new tools to aid clinicians in evaluating mental health, such as natural language processing (NLP) in transcribing video sessions and suggesting clinical notes, as well as scanning patient intake stories and analyzing that information for the likelihood of certain diagnoses. HealthcareITNews

The slow drip-drip-drip of layoffs, folded companies’ asset sales, and company losses that started in 2022 continue, though at a diminished pace compared to consumer companies:

  • Genomics and DNA testing company 23andMe announced layoffs of 9% of staff or 75 people. This will take place by the end of their FY 2024, which ends next 31 March. In a 9 June filing, the company claimed that it would reduce annualized payroll and benefit expenses by $12.8 million, which leads one to wonder about the compensation level of those 75 and from what area they are in. South San Francisco-based 23andMe continues to be money-losing, increasing annual net losses from $217 million to $317 million in the 12 months ending in March, according to its May earnings report despite a 10% revenue gain. 23andMe is yet another ‘cracked SPAC’, having gone that route in 2021 with a Virgin-backed SPAC. Once trading at highs of $12-13 on the NYSE, it closed today at $1.96. However, they don’t have debt, are hanging on to a valuation of $924 million, and their cash position is apparently strong enough to hold it for two years.  Becker’s, SF Chronicle, Yahoo Finance, SimplyWallStreet
  • Another well-financed company, Cityblock Health, is laying off 12%, or 155 staff. Spun off from Sidewalk Labs (Alphabet Health-Google) at the end of 2017, their CEO announced the layoffs in a blog post late last week and effective immediately for those affected. Cityblock serves Medicaid and low-income ‘duals’ with both Medicare and Medicare in value-based care models with a heavy reliance on technology. Their CEO who joined the company in March phrased it as a restructuring, using technology to automate processes, and reducing staff layer. In contrast to others, Cityblock has had no trouble raising funding in the past; in December 2020 they raised $160 million in a December 2020 Series C, plus another $192 million in a Series C extension in March 2021, then a reported $400 million Series D in September 2021 with total raises over $890 million. But their cash burn with high-cost operations in six states (HQ New York, Massachusetts, Indiana, North Carolina, Ohio, and Washington DC) is also likely high. FierceHealthcare
  • Thirty Madison buys assets from bankrupt The Pill Club. The assets? Over 100,000 patient files for $32.3 million. The Pill Club entered Chapter 11 in April after being charged by California authorities of defrauding the state Medicaid program. It paid $18.3 million to settle the charges. But that wasn’t all. According to Mobihealthnews, “the settlement came just days after a state court unsealed a whistleblower complaint against the company in which former nurse practitioners alleged it had defrauded private insurers in at least 38 states. According to a statement from their attorneys regarding the settlement, the whistleblowers would receive approximately $5 million.” The patients covered by The Pill Club’s prescriptions will be converted over to another Thirty Madison brand, Nurx, and offered other services such as behavioral health and dermatology services. The Pill Club raised a lot of money from 2016–$51 million in Series B funding in 2019 and another $41.9 million in 2021. Thirty Madison is a private multi-line of consumer-marketed brands such as Keeps (hair), Picnic (allergies), Cove (migraine), and Facet (psoriasis, eczema) and is at a Series C with a total raise of $209 million. Axios, 
  • Butterfly Network, which some years back developed a hand-held ultrasound device (Butterfly iQ) and entered a crowded field with GE HealthCare’s VScan, Mobisante (apparently defunct), and Philips Lumify, reported Q1 revenue of $15.5 million, flat year-over-year compared to Q1 2022 and which missed analyst estimates (again). Somewhat better news was narrowing last year’s loss to a Q1 loss of $33.5 million which was less than Q1 2022’s loss of $44.5 million. It’s another early SPAC that hasn’t had a great time of it. Since its debut on the NYSE in December 2021, the stock has had the typical drop in altitude from $19.79 to $2.42. It has since expanded to enterprise imaging with Blueprint. Mobihealthnews, Yahoo Finance, Zachs

IBM Watson Health’s stumble and possible fall

July 21, 2021 | by

This Editor hadn’t thought about or seen news about IBM Watson Health in over a year…and likely, neither did you. Granted, our minds have been Otherwise Engaged, but for the company that was supposed to dominate AI and health analytics, it’s notable that TTA’s last two articles mentioning Watson Health was 25 April 2019, on a report that its Drug Discovery unit was being cut back as the latest in a series of executive cutbacks and lawsuits (MD Anderson on a failed oncology initiative), and 14 Feb 2020 on 3M’s lawsuit on unauthorized use of their software.

The New York Times in an investigative piece (may be paywalled or require signup for limited access), brings us up to date on what is happening at IBM Watson, and it’s not bright for Watson Health. IBM, like so many other companies, badly underestimated the sheer screaming complexity of health data. Their executives believed they could translate the big win on the “Jeopardy!” game show in 2011, based on brute computing power, into mastery of healthcare data and translation into massive predictive models. The CEO at the time called it their ‘moon shot’. Big thinkers such as Clayton Christensen chimed in. IBM managers sang its praises to all in healthcare who would listen. This Editor, on a gig at a major health plan in NJ that was ‘thinking big’ at the time and used IBM consultants extensively, in 2012 was able to bring in speakers from Watson for an internal meeting.

But we haven’t been on the moon since 1972 (though probes have visited Mars). Since the big push in 2011-12, it’s been one stumble after another. According to the Times:

  • The bar was set much too high with oncology. Watson researchers knew early on in their research at the University of North Carolina School of Medicine that their genetic data was filled with gaps, complexity, and messiness. The experience was similar with Memorial Sloan-Kettering Cancer Center. The products growing out of the UNC and MSKCC research, Watson for Genomics and Watson for Oncology, were discontinued last year. These were in addition to the MD Anderson Cancer Center initiative, Oncology Expert Advisor for treatment recommendations, that was kicked to the curb [TTA 22 Feb 2017] after $62 million spent. At the same time, IBM’s CEO was proudly announcing at HIMSS17 that they were betting the company on multiple new initiatives. 
  • Watson Health, formed in 2015, bought leading data analytics companies and then didn’t know what to do with them. TTA noted in August 2018 that Phytel, Explorys, and Truven Health Analytics were acquired as market leaders with significant books of business–and then shrank after being ‘bluewashed’. HISTalk, in its review of the Times article, noted that along with Merge Healthcare, IBM spent $4 billion for these companies. IBM’s difficulties in crunching real doctor and physical data were well known in 2018 with revealing articles in IEEE Spectrum and Der Spiegel. 

Six years later, Watson Health has been drastically pared back and reportedly is up for sale. Smaller, nimbler companies have taken over cloud computing and data analytics with AI and machine learning solutions that broke problems down into manageable chunks and business niches.

What’s recoverable from Watson? Basic, crunchy AI. Watson does natural language processing very well, as well as or better than Amazon, Google, and Microsoft. Watson Assistant is used by payers like Anthem to automate customer inquiries. Hardly a moonshot or even clinical decision support. For business, Watson applications automate basic tasks in ‘dishwashing’ areas such as accounting, payments, technology operations, marketing, and customer service. The bottom line is not good for IBM; both areas bring in a reported $1 billion per year but Watson continues to lose money. 

A sobering, mercifully hype-free view of AI in healthcare

October 3, 2018 | by

Way up there on the Peak of Inflated Expectations in the Gartner Hype Cycle is that two-letter creature, AI. Artificial Intelligence has been invoked in multiple tech fields, and Microsoft in the US currently is running 30 second commercials about how AI is “making tomorrow today” but without much explanation as to how.

If AI’s current puffery makes you dizzy, long-time observer of the Healthcare Scene Anne Ziegler’s article in Hospital EMR and EHR might stabilize the whirlies. In direct and brief terms, she classifies the realities of healthcare AI adoption in three areas:

  1. Lack of Transparency. How does AI reach its conclusions in making ‘good decisions’? Sometimes the logic of the conclusion is obvious, but often it is not, and what you get is physician and clinician bypass–and suspicion.
  2. That Old Monkey Wrench Tossed into Existing Processes. It’s taken a long time for organizations to fully integrate their EHR inputs and documentation. Throwing in an AI implementation even in a limited sense may require more adjustments than the outcomes are worth.
  3. It’s Too, Tooooo Much Data. Healthcare organizations do not suffer from a paucity of data. AI feeds on data. Sounds like a good match, doesn’t it. Except that a lot of this data isn’t usable without filtering and mining, and that takes a lot of processing. The future may have more advanced data processing and indexing tech to do that, but right now even natural language processing to identify useful information is rare in the field.

Widespread AI use in healthcare is, despite the IBM Watson Health hype, a long way off. In healthcare, the rubber must meet the road of patient care and clinical practicality to be useful to us with Non-Artificial Intelligence. Problems We Need To Address Before Healthcare AI Becomes A Thing