24/01/2026
Scientific Innovations from 2025
Automated Insulin Delivery Boosts Type 2
Diabetes Control
A GROUNDBREAKING multicentre trial has shown
that automated insulin delivery (AID) systems
can substantially improve glucose control in
adults with insulin-treated Type 2 diabetes.
While AID has previously proven effective in
Type 1 diabetes, high-quality evidence for its
use in Type 2 diabetes has been lacking.1
The 13-week randomised study enrolled 319
participants who were assigned in a 2:1 ratio
to either an AID system or to continue their
usual insulin regimen, with all participants
using continuous glucose monitoring. The
trial’s primary outcome was the change in
HbA1c at Week 13.
Findings revealed a marked improvement in
glycaemic control for those using AID. HbA1c
decreased from 8.2% to 7.3% in the AID group,
compared with a modest reduction from 8.1% to
7.7% in the control group, representing a
significant mean difference of -0.6 percentage
points. Time spent within the target glucose
range (70–180 mg/dL) increased from 48% to 64%
with AID, while the control group showed
virtually no change. Secondary measures of
hyperglycaemia also favoured AID, highlighting
its ability to maintain glucose levels closer
to optimal targets. Hypoglycaemia was rare
across both groups, with only a single severe
event reported in the AID arm.
The results suggest that AID systems can
deliver clinically meaningful improvements
over standard insulin therapy with continuous
glucose monitoring alone. By automating
insulin adjustments based on real-time glucose
readings, these systems reduce the burden on
patients while enhancing safety and efficacy.
As diabetes management increasingly embraces
technology, AID represents a significant step
towards personalised, responsive care for
people with Type 2 diabetes. The trial
provides compelling evidence that automated
systems, once primarily reserved for Type 1
diabetes, may soon play a pivotal role in
optimising outcomes for a broader patient
population.
Alzheimer’s Disease Prediction: A Multi-agent
Large Language Model Framework
A MULTI-AGENT large language model framework
has demonstrated that it can feasibly support
early risk assessment for Alzheimer’s disease
by extracting clinically relevant indicators
from unstructured records and generating more
accurate long horizon predictions than single
model approaches.2
Early detection of Alzheimer’s disease relies
heavily on assessments and biomarker tests
that are expensive, invasive, or impractical
for population-level use. Although structured
electronic health record data have been used
for risk prediction, unstructured clinical
narratives contain valuable information that
is often overlooked. Recent advances in
language models have created opportunities to
analyse free text at scale; however,
challenges persist regarding privacy,
scalability, and the limited reasoning
capacity of single models for complex
diagnostic tasks. These limitations have
prompted interest in collaborative language
model systems that emulate multidisciplinary
clinical reasoning.
To address these issues, researchers developed
CARE-AD, a multi-agent framework for
forecasting Alzheimer’s disease onset using
longitudinal electronic health record notes.
The system assigns specialised agents to
clinical domains including primary care,
neurology, psychiatry, geriatrics, and
psychology. Each agent reviews temporally
ordered narratives and produces a domain-
specific assessment of symptoms and functional
decline. A separate Alzheimer’s specialist
agent integrates these assessments into a
single risk estimate. The architecture is
designed to capture early and often subtle
patterns that may be underrepresented in
structured data, while providing
interpretable, agen-level outputs that
clinicians can examine.
In a retrospective study, using data from the
United States Veterans Health Administration
(VHA), CARE-AD achieved higher accuracy than
baseline single model methods across all
prediction windows. At 10 years prior to the
first diagnosis code, accuracy was 0.53,
compared with 0.26–0.45 for benchmark
approaches. Performance remained strong at
closer intervals, reaching 0.83 at 1 day
before diagnosis. These findings show that
clinically meaningful signals appear in
unstructured narratives well before formal
diagnosis, and that multi-agent modelling
improves detection of such patterns.
The results indicate that multi-agent language
model systems may offer a promising pathway
for earlier identification of Alzheimer’s
disease and could enhance clinical decision
support by providing transparent, domain-
specific reasoning. Future work should include
prospective validation, assessment across
additional populations, and integration with
workflow-compatible clinical tools to
determine how such systems can best support
risk stratification and targeted early
intervention in practice.
Digital Self-Management Programmes Enhance
Asthma Control
DIGITAL health tools are increasingly explored
as scalable strategies to support asthma
self-management, yet robust evidence for their
long-term effectiveness remains limited. In
this pragmatic randomised clinical trial,
investigators evaluated whether a digital
asthma self-management (DASM) programme could
improve symptom control among adults with
asthma compared with usual care.3
The decentralised, open-label study enrolled
901 adults with asthma, who were recruited
remotely and followed for 12 months,
reflecting real-world implementation outside
of traditional clinical settings.
Participants randomised to the DASM
intervention used an app-based platform
incorporating symptom logging, tailored
notifications, wearable device integration,
and self-management tools, while control
participants received usual care. The primary
outcome was change in Asthma Control Test
(ACT) score, a validated patient-reported
measure of asthma control. Secondary outcomes
included engagement with the digital platform
and self-reported medication adherence.
Among participants with uncontrolled asthma at
baseline, the DASM programme was associated
with significantly greater improvement in
asthma control compared with usual care over
12 months. DASM users experienced a mean ACT
improvement of 4.6 points, compared with 1.8
points in the control group, resulting in a
clinically and statistically significant
adjusted between-group difference. These
findings suggest that digital self-management
support can meaningfully enhance symptom
control when added to standard care.
Subgroup analyses revealed that race moderated
the observed treatment effect. While non-
African American participants demonstrated
substantial benefit from DASM, the difference
between intervention and control groups among
African American participants was smaller and
not statistically significant. No significant
moderation was observed by insurance status or
Hispanic ethnicity, indicating broadly similar
effects across these sociodemographic factors.
The reasons underlying differential response
by race remain unclear and warrant further
investigation.
Overall, this trial provides evidence that a
DASM programme can improve long-term symptom
control in adults with asthma, particularly
among those with uncontrolled disease at
baseline. The findings support continued
development and refinement of digital asthma
interventions, with attention to ensuring
equitable effectiveness across diverse
populations.
H-Scan Ultrasound Provides Rapid, Accurate
Kidney Fibrosis Assessment
IN A NEW study, researchers have developed a
quantitative algorithm, called renal H-scan,
that accurately estimates kidney quality at
the time of considering it as a donor organ.
This ensures that donor recipients have the
highest chance of survival by receiving a
fully functional, high-quality kidney.4
Kidney transplantation is currently the
optimal treatment for renal failure. The
current procurement process in the USA
involves carrying out a biopsy to measure
renal fibrotic burden, a critical measure of
irreversible kidney injury. However, these
biopsies come with limitations. Inaccuracies
can be introduced by sampling bias and rapid
sample preparation. Moreover, biopsy carries a
risk of bleeding and requires continuous
access to trained pathology expertise, which
is not always feasible. Importantly, the
highly localised nature of biopsy samples
often fails to represent fibrosis across the
whole kidney, limiting their predictive value
for post-transplant renal function.
In a first-in-human study, researchers have
developed renal H-scans, which can be
integrated into standard ultrasound workflows
as another option to measure the fibrotic
burden quickly and non-invasively. The H-scan
algorithm has been tested in preclinical
animal models and human transplant kidneys,
offering the ability to assess fibrosis across
the entire organ rather than a small,
localised sample. Unlike traditional biopsy,
H-scan estimates of whole-kidney fibrosis were
found to correlate closely with renal function
following transplantation, suggesting superior
predictive potential for post-transplant
outcomes.
This innovation has several important
implications. By providing an accurate, non-
invasive assessment of kidney quality, H-scans
could reduce reliance on biopsies, minimising
procedural risks and the need for round-the-
clock pathology support. The technique is
straightforward to implement within existing
ultrasound workflows, making it accessible and
scalable for clinical practice. Additionally,
accurate whole-kidney fibrosis quantification
could improve organ allocation decisions and
help ensure that transplant recipients receive
kidneys with the highest likelihood of
supporting long-term function.
Overall, the renal H-scan represents a novel,
practical, and clinically impactful tool for
improving the assessment of donor kidney
quality. Its adoption could enhance patient
safety, optimise transplant outcomes, and
address critical limitations associated with
biopsy-based fibrosis assessment.
Deployment of a Digital Clinical Alert
Navigation System in Integrated Care
A RESEARCH study has demonstrated that natural
language processing-driven clinical alerts and
navigation tools can improve high acuity
detection, support appropriate care routing,
and contribute to an enhanced patient
experience within an integrated, value-based
care model.5
Patient portals have become central to digital
self-service in healthcare, yet ensuring that
such systems reliably guide patients towards
clinically appropriate pathways remains a
major challenge. In response to growing
concerns related to self-scheduling, the
Southern California Permanente Medical Group
(Pasadena, USA) established the Virtual Safety
Net, a natural language processing-enabled
ecosystem that successfully identified time
sensitive risks, but could not be readily
scaled. These results highlighted the need for
a more comprehensive, adaptable system,
capable of embedding safety, navigation
accuracy, and user centric design into routine
digital interactions. Therefore, the Kaiser
Permanente Intelligent Navigator (KPIN; Kaiser
Permanente, Oakland, California, USA) was
developed to integrate clinical alerts and
navigation tools into the appointment booking
workflow.
The KPIN system processes input through a
multilingual natural language pipeline that
incorporates large language models and custom
transformer architectures. These models
evaluate symptoms, identify high acuity
presentations, retrieve relevant demographic
and clinical guideline data, and generate a
curated set of appropriate care options.
Outputs include video or phone visits,
asynchronous messaging, or other modalities
consistent with clinical standards. Encounters
conclude when patients select an offering,
whereas abandonment represents incomplete
interaction.
KPIN demonstrated strong performance in
detecting high acuity symptoms, achieving
96.0% accuracy (95% CI: 93.7–98.0%), with
97.5% precision (95% CI: 95.8–99.0%), and a
recall of 96.0% (95% CI: 93.8–97.9%).
Similarly, the clinical navigation models
achieved 81.9% accuracy (95% CI: 80.0–83.6%),
with a corresponding precision of 85.6% (95%
CI: 84.0–87.2%), a recall of 81.9% (95% CI:
80.1–83.7%), and an F1-Score of 82.8% (95% CI:
81.1–84.5%). Additional metrics showed that
KPIN’s adjusted successful booking rate was
53.68%, with an abandonment rate of 2.94%
(interquartile range: 2.77–3.11%), aligning
with patient survey results showing an 8.63
percentage point increase for positive
sentiment.
These findings indicate that an integrated
digital navigation system can enhance clinical
safety, streamline access, and support value-
based care by directing patients to suitable
modalities with greater accuracy. Future work
should focus on establishing pre-
implementation baselines, expanding
conversational capabilities, and evaluating
real-world clinical outcomes to determine how
such systems can best complement clinician
workflows and improve timely access to
appropriate care.
New Microneedle Device Tracks Lactate in
Pregnancy
RESEARCHERS in Liverpool, UK, have
successfully tested a new microneedle device
that can continuously monitor lactate levels
in pregnant women, a development that could
transform maternity care by enabling earlier
detection of complications such as sepsis
during labour.6
The pilot study, conducted at a clinical
research facility in a Liverpool hospital,
involved seven healthy, pregnant volunteers
over the age of 18 years. All participants had
uncomplicated pregnancies and were accustomed
to light exercise. Each woman wore a small
microneedle patch designed to measure lactate
levels in the skin’s interstitial fluid, while
simultaneous blood samples were taken to
compare results.
During a 30-minute session of gentle exercise
followed by rest, the device tracked changes
in lactate levels continuously. Researchers
found that interstitial lactate closely
mirrored venous lactate trends, suggesting
that the new technology could offer a
reliable, non-invasive alternative to
traditional blood testing. Participants
reported minimal discomfort, with average pain
and discomfort scores of just 0.43 and 0.14
out of 10, respectively.
As stressed by the researchers, little is
known about lactate levels during labour, with
some expectation of them rising due to
anaerobic respiration during physical
exertion. Elevated lactate levels,
particularly above 2 mM, can prompt clinicians
to consider sepsis, a potentially life-
threatening condition requiring immediate
treatment. Current lactate measurements
provide only snapshots in time, but the new
device offers continuous, real-time insight
into a patient’s condition.
Although this proof-of-concept study involved
only a small group of healthy volunteers,
researchers say the findings point to
significant future applications. Continuous
lactate monitoring during labour could help
clinicians identify early signs of distress or
infection, enabling swift intervention while
also avoiding unnecessary use of antibiotics.
Virtual Reality Meditation Eases Depression
and Anxiety
A NEW clinical study has found that immersive
virtual reality meditation (IVRM) may offer a
promising, non-pharmacological option for
reducing symptoms of major depressive disorder
and generalised anxiety disorder among
hospital inpatients. The research adds to
growing interest in technology-enabled
mindfulness therapies, which aim to increase
access to evidence-based psychological
support.7
Mindfulness-based cognitive therapy has long
been recognised as an effective complement to
traditional treatment for depression and
anxiety. However, questions remain about
whether technology-assisted approaches can
achieve similar clinical benefits,
particularly within inpatient settings where
symptom severity is often higher. This study
explored whether IVRM, an immersive adaptation
of mindfulness-based cognitive therapy, could
improve emotional regulation and, in turn,
reduce depressive and anxiety symptoms.
The 10-week, single-arm clinical trial
involved 26 participants at a community
hospital behavioural health unit, each
diagnosed with major depressive disorder and
generalised anxiety disorder. Patients engaged
in IVRM sessions three times per week, with
outcomes monitored using depression and
anxiety assessments alongside ECG-based
measures of emotional regulation. The study
focused particularly on the Coherence
Achievement Score (CAS), a physiological
indicator reflecting the stability and balance
of emotional responses.
Researchers analysed associations between CAS,
symptom changes, and relevant covariates using
a generalised estimating equation model.
Results showed a clear relationship:
improvements in emotional regulation following
IVRM sessions were associated with meaningful
reductions in both depression and anxiety. The
findings support the idea that virtual reality
may not only enhance engagement with
mindfulness practices, but also strengthen the
emotional regulatory mechanisms underpinning
therapeutic improvement.
While the study is limited by its small sample
size and single-arm design, it contributes
valuable insights into the feasibility and
clinical potential of IVRM in inpatient mental
health care. As interest grows in innovative,
accessible mental-health interventions,
immersive virtual reality may become an
important tool for hospitals seeking
complementary treatments that reduce symptom
burden without adding pharmacological side-
effects.
Carbon Emissions of Virtual Wards Compared
with Inpatient Care
DIGITALLY enabled models of care are
increasingly central to healthcare delivery
within the NHS, offering opportunities to
expand capacity while supporting
sustainability goals. As virtual wards are
scaled nationally, understanding their
environmental impact is essential to aligning
innovation with the NHS commitment to
achieving net zero carbon emissions. This
study evaluated the carbon footprint of a
virtual ward pathway for patients with acute
respiratory infections and frailty in a large
acute hospital trust, using a retrospective
cohort analysis.8
Carbon emissions associated with virtual ward
care were compared with those generated by
traditional inpatient bed days. The analysis
demonstrated that virtual wards were
associated with a substantially lower carbon
impact, largely driven by reductions in
hospital-based resource use and patient
travel. These findings support the premise
that home-based digital care pathways can
contribute meaningfully to healthcare
decarbonisation when implemented at scale.
The study also compared two approaches to
carbon accounting: a manual data audit and an
automated business intelligence extraction.
While automated methods offered efficiency and
scalability, manual audits provided more
granular insights into emissions related to
community care and out-of-hospital activity.
This discrepancy highlights current
limitations in routinely collected datasets
when assessing the full environmental impact
of digital care models.
Overall, the findings suggest that virtual
wards represent a lower-carbon alternative to
inpatient care for selected patient groups.
However, accurate measurements of their
environmental benefit depend on improved data
capture beyond the hospital setting. Enhancing
automated systems to better reflect
community-based care will be critical for
robust carbon evaluation as digitally enabled
services continue to expand across the NHS.
Wearable Sweat Sensor Enables Remote
Monitoring in Cystic Fibrosis
SWEAT chloride concentration is a key
biomarker in cystic fibrosis (CF),
traditionally measured using pilocarpine-
induced sweat collection and chloridometry.
While this method remains the diagnostic gold
standard, it is impractical for frequent or
remote monitoring. This study evaluated a
wearable skin-interfaced microfluidic ‘CF
Patch’ (Epicore Biosystems, Cambridge,
Massachusetts, USA), paired with smartphone-
based image analysis, as a tool for real-time
and remote assessment of sweat biomarkers in
adults with CF.9
Clinical studies compared CF Patch
measurements with standard chloridometry under
both pilocarpine-induced and exercise-induced
sweating conditions. In laboratory settings,
the CF Patch demonstrated strong correlations
with sweat chloride values obtained via
pilocarpine-based chloridometry, regardless of
whether sweat was induced pharmacologically or
through exercise. These findings support the
accuracy of the device for quantifying sweat
chloride and sweat rate in controlled
environments.
The feasibility of remote monitoring was also
assessed. In healthy volunteers, exercise-
induced sweat chloride measurements collected
remotely using the CF Patch were strongly
correlated with in-laboratory CF Patch
measurements. In contrast, correlations were
weaker in people with CF, particularly among
those receiving cystic fibrosis transmembrane
conductance regulator (CFTR) modulator
therapy. Importantly, individuals with CF on
modulators showed greater day-to-day
variability in sweat chloride levels compared
with healthy volunteers. This variability
highlights a limitation of relying on single,
in-laboratory chloridometry measurements to
assess CFTR modulator efficacy and
pharmacodynamics.
Overall, the findings indicate that the CF
Patch enables the serial, non-invasive
measurement of sweat chloride in both
laboratory and remote settings. While the
device is not intended to replace
pilocarpine-induced chloridometry for
diagnostic purposes, it shows promise as a
remote disease management tool. By allowing
repeated measurements over time, the CF Patch
may provide more nuanced insights into
treatment response and medication
effectiveness in adults with CF.
Leveraging Large Language Models to
Personalise Therapy in Rare Gynaecologic
Malignancies
RARE gynaecological tumours (RGT) remain a
difficult concept in oncology due to their low
incidence, pronounced heterogeneity, and
absence of robust, evidence-based clinical
guidelines. These challenges often result in
delayed diagnosis, limited therapeutic
direction, and poor clinical outcomes for
affected patients. While molecular tumour
boards provide a promising route towards
individualised therapy by leveraging
biomarker-guided decision-making, their impact
is frequently constrained by the fragmented,
unstructured nature of clinical and molecular
data. Manual data curation slows the process,
limits scalability, and increases the risk of
overlooking relevant therapeutic
opportunities.10
This study evaluates whether large language
models (LLM) can overcome these barriers by
enabling the construction of digital twins,
computational patient replicas that integrate
detailed clinical, pathological, and biomarker
information. The authors developed a proof-
of-concept LLM-enabled digital twin framework
combining data from institutional cases and
published case reports (21 patients), with a
comprehensive literature dataset derived from
655 peer-reviewed publications. The system was
applied to metastatic uterine carcinosarcoma,
a highly aggressive and understudied RGT with
limited treatment consensus.
The LLM-powered digital twins demonstrated the
capability to synthesise multimodal data
streams, harmonise terminology, and generate
personalised treatment recommendations.
Importantly, the digital twin approach
identified potential therapeutic strategies
not captured through traditional, single-
source analyses, highlighting the value of
cross-referencing diverse datasets and
expanding the evidence base for rare, complex
cancers. By modelling possible clinical
trajectories, the system also offers a means
to anticipate disease progression and refine
therapy choices over time.
This work highlights a significant conceptual
shift in oncology: moving from an organ-based
categorisation of tumours to a biology-driven,
biomarker-centric framework. LLM-enabled
digital twins have the potential to accelerate
precision oncology for RGTs by improving data
accessibility, reducing manual workflow
burden, and more effectively linking patients
to targeted therapies. As digital twin
technologies mature, they may help clinical
teams overcome the inherent evidence gaps in
rare tumour research, ultimately supporting
more individualised care pathways and
improving outcomes for patients with rare
gynaecological malignancies.
