This study develops a generative clinical LLM using 277 billion words of text and up to 20 billion parameters. The model improves biomedical natural language processing, generates synthetic clinical text, and passed Turing test in writing clinical notes.
In this perspective,
@Berci
&
@EricTopol
discuss the regulation of GPT-4 & generative
#ArtificialIntelligence
in medicine... balancing the exciting + transformative potential, but ensuring safety, maintaining ethical standards, & protecting privacy.
Foundation models (FMs) such as
#ChatGPT
have the potential to revolutionize healthcare. But what's hype and what's real?
This review from a team in
@StanfordMed
includes 84 clinical FMs + proposes an evaluation framework better suited to assess value.
This study developed an open-source tool using a LLM to extract important medical information from clinical text, focusing on decompensated
#LiverCirrhosis
.
The tool identified liver cirrhosis from free text with 100% sensitivity and 96% specificity, and showed strong results
Your weekend read editorial: Discussing the importance of shifting the focus towards clinically relevant outcomes, when adopting
#ArtificialIntelligence
tools, the ecosystem required for AI to succeed in health, & the human aspect of
#healthcare
.
This article introduces ETHOS, an AI tool that uses advanced
#MachineLearning
to predict future health outcomes based on patient records, without needing labelled data or model tuning. This tool can simulate different treatment options, helping improve patient care & reduce
Are large language models (LLMs) safe for use in medicine? In this study led by
@OmiyeTofunmi
&
@RoxanaDaneshjou
, the authors found that four different LLMs had outputs that perpetuated false race-based medicine.
(1/4) Can AI help surgeons improve their skills?
We are delighted to share a trio of articles published across
@NaturePortfolio
covering AI-based assessment of surgery skills, authored by
@DaniKiyasseh
and colleagues.
All articles are fully open access 🔓
A curated database of
@US_FDA
-cleared
#ArtificialIntelligence
products for medical imaging indicates a high prevalence of triage systems, variability in basic output characteristics, and often limited
#explainability
of AI predictions.
A comprehensive review of natural language processing over the past 15 years from a team in the UK. NLP budgets from research funding now are 80x that from 2007-10, yet there is still much to do in this field to realise its potential.
Of 4 collaboration strategies to deploy
#AI
for the diagnosis of ARDS from chest XR, the most accurate is to allow AI review the XR first & defer to the physician if uncertain. 79% of cases were decided by AI - significantly reducing physician workload.
"Data have become the most valuable commodity in health care, but questions remain about whether there will be an
#AI
'revolution' or 'evolution' in health".
This perspective article discusses action areas & recommendations to achieve AI's full potential.
Generative
#ArtificialIntelligence
can create synthetic patient data that closely mimics disease biology, treatment response & long-term survival compared to real clinical trial cohorts. Synthetic patient data may augment control groups in trial designs.
Gestational age estimation has an error of +/- 2 weeks in late
#pregnancy
. The application of this
#machinelearning
model using only ultrasound image analysis reduces this error to 3 + 4 days in the 2nd & 3rd trimesters representing a major step forward.
With advancements in data science &
#AI
, the concept of
#DigitalTwin
for Health holds immense potential.
This scoping review explores its promising applications, paving the way for a collaborative global effort in healthcare innovation.
What can we learn from how you type on your phone?
This study shows that it’s possible to track and predict fluctuations in people’s mental health status based on how they type on their phone with nearly 95% accuracy.
Introducing
@drjessilyn
, Jessilyn joined us as an Associate Editor in March 2021 and is one of our longest-serving Associate Editors. Jessilyn is highly experienced in assessing manuscripts focused on digital biomarkers and machine learning.
This Perspective, written by one of our previous Editors-in-Chief,
@EricTopol
, and
@Berci
continues to attract readers and was our 🏆 top
@altmetric
article for July 2023.
A Perspective in
@npjDigitalMed
argues that regulatory oversight should assure medical professionals and patients can use large language models and AI without causing harm or compromising their data or privacy and offers recommendations to regulators.
New this week in
@Nature
: A new
#AI
tool designs
#mRNAvaccines
that have greater shelf stability and trigger a larger antibody response in mice than conventionally designed vaccines.
Smartphones &
#wearables
transform health monitoring. However, bias in digital health studies and algorithms arises from underrepresented datasets.
This
@DukeU
team explore device ownership, reasons for use, & willingness to share data for research.
Using a phone app to monitor wounds for infection after surgery... In this multi-centre pilot, 200 patients felt the app improved almost every aspect of their wound care after surgery including usefulness, ease of use, reliability of tech, & satisfaction.
Large language models show promise for identifying social determinants of health in clinic notes, with smaller, fine-tuned models showing more robust performance & less potential bias than ChatGPT-family models.
A single approach that can recognise walking with a variety of body-worn devices regardless of walking style, sensor location and configuration is presented in this paper to estimate walking time, cadence and step count.
Large language models can rapidly phenotype patients with no manual annotation. This study reports on the evaluation of a publicly available LLM for zero-shot phenotyping using clinical notes with a proof-of-concept in postpartum hemorrhage.
We would like to welcome Prof.
@girish_nadkarni
from
@IcahnMountSinai
to our Associate Editor team. Girish is specialized in AI and emerging technologies in medicine.
Our June Reviewer of the Month is
@kdpsinghlab
! On peer review, Karandeep comments: "With all the rapid innovations happening right now in the space of AI & large language models, it's more important than ever to separate hype from reality."
Can synthetic data improve AI models?
In this paper, a generative model for synthesizing mixed-type timeseries health records, including physiological signals, lab results, & medication uses, improved the downstream applications of AI in critical care.
This team from
@cerebral
built a
#MachineLearning
system to identify chat messages from patients experiencing
#mentalhealth
crises.
When deployed in a national telehealth network it reduced response times to patients in crisis from 10 hours to <10 mins.
Introducing
@JohnTorousMD
, John joined us as an Associate Editor in November 2021. John is highly experienced in assessing manuscripts focused on digital psychiatry and mental health.
Can a novel
#DeepLearning
algorithm to predict
#sepsis
in the emergency department improve care?
Researchers from
@UCSDHealth
show proper implementation of the model can save lives of sepsis patients and improve quality metrics.
This perspective proposes a framework to improve computational phenotyping by enhancing accessibility & usefulness. The authors present an example focussing on easy adoption, cross-task reusability, and clinical phenotyping algorithm development support
We are thrilled to welcome Professor
@adamdunn
from
@Sydney_Uni
to our Associate Editor team! Adam brings a wealth of public health & clinical informatics expertise, with a particular interest in online misinformation & health behaviors.
Mitigating the 'black box' of AI: LLMs can imitate diagnostic reasoning strategies when solving clinical cases & provide an interpretable means to assess if the generated answer is true/false based on the diagnostic reasoning's factual & logical accuracy.
📢 Open for submissions, our new collection, Clinical Applications of AI in Mental Health Care. This collection is in collaboration with
@npjMentalHealth
& is Guest Edited by Eric Kuhn,
@MauriceMulvenna
, Raymond Bond, &
@JohnTorousMD
.
Germany's DiGA program was the first to implement
#DigitalTherapeutics
on a large scale. Three years later, increased usage and accumulated experience have prompted regulatory changes that expand the program's scope.
This umbrella review summarizes 48
#systematicreviews
on effectiveness of app-based health interventions within patient populations published from 2013-23.
Find out how skewed the evidence is towards some indications & where quality remains suboptimal:
Claims about how
#AI
will increase efficiency are prominent but flawed, because they are based on studies of accuracy.
Accuracy ≠ efficiency...
We must remain conscious that an accurate model does not guarantee efficiency gains & workload reduction.
A novel wearable-sensor dataset of healthcare workers from
@sujnagaraj
et al. reveals high levels of inter- & intra-individual heterogeneity in stress - an important step towards characterizing
#stress
in individuals.
Using datasets of EHR +
#echofirst
,
@David_Ouyang
et al. found confounders mediate AI prediction of demographics. Unaccounted confounders provide a substantial amount of the information in “predicting race” and performance drops when balancing covariates.
Breaking news 💥 Our Guest Editor Raymond Bond announces the launch of our new collection call for papers, Clinical applications of AI in mental health care. This is a collaboration with new launch
@Nature_NPJ
journal
@npjMentalHealth
.
Advancing EHR-based deep phenotyping... An algorithm for mapping clinical Observational Medical Outcomes Partnership (OMOP) vocabularies to
@OBOFoundry
ontologies to help systematically identify undiagnosed patients who might benefit from genetic testing.
A
#deeplearning
model, the Stanford Estimator of Electrocardiogram Risk, predicts long-term risk of
#cardiovascular
mortality and disease from a single ECG with performance better than current standard risk screening tools.
Scaling AI for surgical guidance to operating rooms globally remains a major challenge, especially resource-limited settings. These authors describe a novel equipment-agnostic
#MachineLearning
pipeline to enable real-time deployment to any edge device.
Health data justice: "the study and use of health-related data in ways that aim to redress the exclusions of structurally marginalized communities from systems of health care and public health…" An interesting piece from
@jayshaw29
&
@sharifasekalala
.
Artificial intelligence for diagnosing psychological & neuro disorders is frequently studied in labs but rarely implemented clinically. This review outlines the biological, technical & institutional barriers to making diagnostic AI a reality in hospitals.
Deep learning of ECGs can help identify patients with hidden
#atrialfibrillation
better than predictions based on clinical risk factors & echo measurements.
This may offer additional opportunities to guide patient screening & support early intervention.
Our most recent News & Views article examines the ethical tension between "selective deployment," where AI is used only for subgroups it performs well in, and "equitable deployment," which aims for fair performance across all groups.
This article emphasizes the need for better
An
#RCT
led by
@JRGolbus
&
@bnallamo
from
@umichCVC
with 220 participants showed that contextually-tailored text messages from wearable devices may improve physical activity for some cardiac rehab participants but displayed no sig. difference at 6 months.
Surgeons with different experience levels use different surgical gestures in robotic-assisted surgery. These gestures were predictive of 1-year postop function, suggesting gestures could objectively measure surgical performance + outcomes.
Exploring evaluation metrics for health
#Chatbots
. What is the current norm? What are their weaknesses?
This perspective also introduces a set of user-centered evaluation metrics, grouped into accuracy, trustworthiness, empathy, & computing performance
Many
#MachineLearning
& AI algorithms struggle to generalize, preventing responsible deployment.
@GSK
AI/ML and
@CC4AIM
use a real-world case study to look at how selective predictions might be implemented technically & explore ethical implications.
Studies of
#AI
in medicine may overestimate accuracy to new situations by 20%.
Models learn hidden biases in X-rays, ECGs, clinical notes etc. instead of real disease markers. This study proposes a solution for inc. reliability in reporting of results.
Christmas Catch-Up: Using a
#digitalhealth
platform to capture patient reported outcomes every 3 days after cardiac surgery showed that patient perception of recovery varies substantially, offering more granular info than conventional outcome metrics.
Multicenter EHR data from >234k adult non-cardiac surgical patients is used to predict maximum pain after surgery.
This
#MachineLearning
model prospectively outperformed clinicians, & can identify comorbidities that drive postoperative pain.
Factors inc. sociodemographics, depression severity, phone status & app usage behavior could be linked to long-term participant retention in remote mobile health studies... Findings from a large European multinational remote digital study for
#depression
.
Building an AI-based 'clock' of physical activity & sleep during pregnancy from >1000 participants' wearable data. This model uses just a smartwatch to predict gestational age. Deviations from this 'clock' are very strongly associated with preterm birth.
📢 Announcing our new collection🌍Harnessing digital health technologies to tackle climate change and promote human health🌍
Guest Edited by
@JRedHeart
@KariceHyun
&
@GeorgiaKatec
. Deadline to submit 30 April 2024.
More info🔗
What are some educational initiatives that you know of designed to educate medical or CS/Eng students about digital health? We’re starting a series to amplify such efforts. Tag them below!
Another for our
#digitalsurgery
readers… This review explores current
#computervision
techniques applied to minimally invasive surgery, the clinical applications and discusses the challenges to wider adoption.
New article! Read Collaborative strategies for deploying AI-based physician decision support systems: challenges and deployment approaches in npj Digital Med
Large language models (
#LLMs
) can effectively summarize medical evidence, but proprietary LLMs often lack transparency and create vendor dependency.
This study showed that fine-tuning open-source LLMs significantly improves their performance, bringing some models, like LongT5,
We love nothing more than a good author thread on their paper!
Here is
@JoschkaHalt
describing his paper with
@RobRanisch
discussing the ethical landscape surrounding the current deployment of LLMs in medicine and healthcare.
1/ How are
#ChatGPT
and other
#LLMs
used in
#Medicine
? And what are the ethical implications?
We say: It’s a social experiment unfolding right in front of us!
Check out
@robranisch
‘s and my new paper in
@npjDigitalMed
Let me give you a quick tour
Measures derived from
@fitbit
+ patient characteristics are used to train
#MachineLearning
models to detect abnormal recovery in children after appendectomy, with results supporting the use of remote monitoring for early detection of postop complications.
This perspective provides an overview of the misalignments of current
#ArtificialIntelligence
fairness research, with discussion on the practical clinical concerns & the obstacles to AI fairness adaptation.
#AIfairness
When adopting
#AI
in the real world, the trade-off between performance & cost-effectiveness needs examination.
This case study on real-world
#diabetic
retinopathy screening in China suggests the most accurate AI may not be the most cost-effective.
Smartphone typing dynamics and accelerometery was predictive of motivation levels & the ability to feel pleasure in suicidal ideation for people assigned female sex at birth in this study from
@LoranKnol
et al.
#mentalhealth
New article! Read Applying machine learning to consumer wearable data for the early detection of complications after pediatric appendectomy in npj Digital Med
A large language model fine-tuned on clinical notes to enhance Diagnosis-Related Group (DRG) assignment - a system to classify hospital cases.
This model showed superior performance compared to prior models in the task of DRG prediction.
Digital health interventions vary widely and control groups used to test the effects of these interventions are even more variable. This perspective proposes names for 11 commonly used control conditions & discusses the scientific implications of each.
Clinicians perceive
#MachineLearning
as an augmentation rather than substitute to their clinical practice. This qualitative analysis by
@KatharineEHenry
,
@suchisaria
et al. explored opinions of clinicians working with a ML-based sepsis warning system...
It's time for another weekend read editorial...
This week, exploring
#GenAI
with Marium Raza,
@KaushikVenkates
&
@jkvedar
. Building on a recent framework from
@MichaelWornow
et al.
this article includes an evaluation checklist for GenAI applied to EHRs.
Using computer vision to transform smartphone videos into highly accurate tools for tremor analysis to rival gold-standard equipment...
These digital biomarkers provide added value for assessing DBS outcomes in patients with essential tremor.
Watch this clip to discover why
@AjhungMD
, a pioneer in medical AI 🌟, chose npj Digital Medicine to publish his research on bias mitigation, surgical AI, and more:
A
#MachineLearning
pipeline to help identify skin tone representation in medical education textbooks.
Notable findings that dermatology textbooks are biased with brown & black skin tone (Fitzpatrick V-VI) images constituting only 10.5% of all skin images
This study on
#AI
implementation finds that causal inference models, contextualized explanations, and ambidexterity between exploration versus exploitation mindsets can help bridge clinicians’ needs and developers' goals to solve the XAI conundrum.
We have 3⃣ exciting open collections welcoming submissions, led by expert Guest Editors.
📌 Digital Technologies for Parkinson's Disease
📌 Clinical applications of AI in mental health care
📌 Regulated digital medical products
Find out how to submit:
Data from wearables can characterize daily physical activity and its change over time in persons with Amyotrophic Lateral Sclerosis, and were significantly associated with clinical patient reported outcome measures… a potential future digital endpoint?