Life Support Center Digital Health and AI in Critical Care LabLife Support Center, Hacettepe University

We keep the physiology that bedside monitors discard, and build clinical AI that holds up to scrutiny.

The Digital Health and Artificial Intelligence in Critical Care Laboratory works inside the pediatric intensive care units of Hacettepe University in Ankara. It is directed by Izzet Turkalp Akbasli, MD, within the Life Support Center founded by Prof. Benan Bayrakci.

II ART Pleth CO₂

Simulated signals. In our units the same streams come from GE CARESCAPE B850 monitors and ventilators through the hospital's Kafka message broker.

A data science lab next to the beds it studies

Founded in 2021 within the Hacettepe University Life Support Practice and Research Center, the lab joins second-by-second physiologic data with the engineering needed to keep it, govern it and learn from it.

Our clinicians are also the people who write the pipelines. We develop machine learning for continuous physiologic telemetry, for acute kidney injury and low cardiac output syndrome after congenital heart surgery, and for extracorporeal therapies such as extracorporeal membrane oxygenation (ECMO) and continuous kidney replacement therapy (CKRT).

A second line asks where clinical AI fails: models that lose accuracy when they move between hospitals, language models working in Turkish records, and vision-language models that can be steered by text inside an image. Our multicenter kidney injury study, for example, reports how widely external performance varied between centers, because that variation decides whether a model is safe to deploy.

The lab is one of the Life Support Center's laboratories, alongside xenotransplantation, hybrid membranes and extracorporeal circuit research, and works with the same clinical teams and the same weekly research board.

1985
Türkiye's first pediatric ICU opened here
23 beds
15 general and 8 cardiac intensive care beds
~800
admissions per year, national quaternary referral center
2007
Türkiye's first ECMO program; about 400 children supported
600+
ECMO days per year
0.28
standardized mortality ratio

NAFIS

Critical Patient Integrated Data and Research Platform

Every monitored child in our units produces physiologic data at second-level resolution. The hospital network already carries these streams, waveform packets included. The hospital information system keeps only minute-level values, and waveforms are deleted from monitor memory shortly after they are drawn. The data is produced and transported, but not recorded. NAFIS records it.

NAFIS listens to the existing device stream without writing anything back, joins it with laboratory, medication, procedure and note data already held by the hospital, and keeps the result inside the hospital network under parental consent and Hacettepe University's institutional data governance policy. It needs no new monitor integration and no additional integration cost. To our knowledge, no pediatric critical care data bank of this kind exists in Türkiye; the comparable resources are WAVES at Stanford, the Texas Children's real-time platform, CHU Sainte-Justine, PIC, MIMIC-IV, HiRID and AmsterdamUMCdb.

NAFIS is a data center, not a laboratory. Our lab designs and runs its architecture and supplies the analytic methods for the studies that use it.

  1. 1Sources

    • Bedside monitors: heart rate, invasive and non-invasive pressures, SpO₂, respiratory rate, etCO₂, temperature, central venous pressure; ECG, plethysmography and pressure waveforms
    • Mechanical ventilators: mode, settings, measured pressures, volumes and flows, alarms
    • Hospital information system: laboratory and microbiology, medications and blood products, procedures, nursing records, severity scores, reports and notes
    • ECMO, CKRT and plasma exchange consoles: structured bedside entry by the physician on duty

  2. 2Ingestion

    A read-only Kafka consumer gathers short segments, writes each one to disk, records it in the catalog, and only then advances its offset.

    The broker keeps 72 hours of messages, so a service outage or hardware failure is replayed without loss. Hospital records arrive through incremental, read-only SQL with a dedicated service account.

  3. 3Storage

    Waveform archive
    Compressed Apache Parquet partitioned by device, data type and date. Device ID, timestamp and values only, never a name or record number.
    Catalog and clinical tables
    PostgreSQL, with constraints that block overlapping time intervals for a device.
    Identity vault
    A separate, restricted schema that links consented children to their random IDs.

  4. 4Research access

    An approved study receives only the variables and time windows it asked for, keyed by random IDs such as NAFIS-004217.

    Analyses run on the in-network server. Only aggregate results leave, after the principal investigator approves them.

Safeguards

  • Read-only access to every source. Devices and hospital systems are never written to or reconfigured.
  • Intranet only. The portal accepts connections from hospital IP ranges, and outbound traffic is closed at the firewall.
  • Disks and backups are encrypted. Master keys are split with Shamir's secret sharing into five shares, any three of which rebuild them: no single person can decrypt, and losing two shares loses nothing.
  • Encrypted incremental backups twice a day to two NAS units at separate sites, immutable snapshots, and a restore test every three months.
  • Personal accounts, role-based permissions enforced on the server, and an immutable audit log of every access, query and export.

Consent

  1. From admission, device data are stored against bed and time only, linked to no child and visible to no researcher.
  2. When the child is stable, a pediatric critical care fellow informs the family and asks for consent, aiming for the first 72 hours. Children aged 7 and older with decision capacity give their own assent on age-appropriate forms.
  3. With consent, the child receives a random ID and the stay moves into the research layer under that ID.
  4. Without consent, the pre-record is deleted unopened by monthly board decision. Bereaved families are not approached; those deletions are counted to monitor selection bias.

Scale and status

≥ 99%
target capture of broker messages
~2 GB per day
about 3.7 TB over the first five years
~3,200 admissions
expected by 31 December 2031, assuming 80% consent
20 years
pseudonymized retention, then anonymization following the MIMIC-IV method: ID replacement, per-patient date shifting, free-text de-identification, key destruction
BAP 23145
Hacettepe University infrastructure project funding the main server and storage (4.8 million TL)
Ethics
Prospective protocol for the Hacettepe University Health Sciences Research Ethics Board; collection starts only after approval

The model travels to the data. The data stays.

Patient-level data never leave Hacettepe. When an external group wants to work with NAFIS, its model is trained or validated inside the hospital network after research board and ethics approval. Only model weights and aggregate performance metrics leave, and the weights are first tested for memorization and membership inference.

What the data bank is built to support
ModalityExamples of future studies (each needs its own ethics approval)
Continuous vital signs and waveformsPredicting cardiac arrest, emergency intubation and severe deterioration hours ahead; early recognition of shock and low cardiac output syndrome; phenotypes from physiologic trends
Ventilator signalsAutomatic detection of patient-ventilator asynchrony; adherence to lung-protective ventilation; extubation readiness; the course of pediatric ARDS
High-resolution ECGElectrophysiologic patterns before arrhythmia; rhythm disturbances after cardiac surgery and ECMO; heart rate variability and clinical course
Clinical records and textPlacing physiologic models in clinical context; drug exposure and physiologic response; structured extraction from de-identified notes and reports
Extracorporeal circuitsAll of the above in children on ECMO, CKRT and therapeutic plasma exchange

Research lines and projects

Each line already has peer-reviewed output or working software. NAFIS gives all of them a prospective, high-resolution source at home. Colored markers show which NAFIS data modalities a line draws on.

Deterioration, physiology and outcomes in the PICU

Predicting which children will deteriorate after cardiac surgery, from routine and high-frequency data, and describing outcomes in our own unit.

  • Hemodynamics
  • Waveforms
  • Clinical records
  • Forecasting low cardiac output in neonatesLightGBM models forecast low cardiac output syndrome 2 to 12 hours ahead from hourly multivariate data in the first 48 postoperative hours (AUROC 0.91 to 0.98), with SHAP explanations.PublishedCritical Care Explorations, 2025
  • Low cardiac output risk at ICU arrivalA parsimonious model from preoperative and intraoperative variables (AUROC 0.88) for early risk stratification after congenital heart surgery.PublishedCardiology in the Young, 2026
  • Methylene blue in vasoplegiaHigh-fidelity Sickbay data, causal inference and cluster analysis identified three hemodynamic response phenotypes in pediatric cardiac intensive care.PublishedPediatric Cardiology, 2026
  • Red blood cell transfusion in the PICUPatterns and predictors in 2,011 children: age, nutritional status and baseline anemia shaped transfusion decisions beyond the hemoglobin value.PublishedNigerian Journal of Clinical Practice, 2026
  • Artificial intelligence in pediatric cardiac intensive careReview of clinical applications, implementation barriers and future directions.PublishedCurrent Treatment Options in Pediatrics, 2026
  • Order of low cardiac output signs in neonates: preregistered time-series causal discovery (PCMCI+) on hourly postoperative data. In revision
  • Machine learning prediction models in pediatric cardiac intensive care: systematic review with PROBAST risk-of-bias appraisal. Under review
  • Critical pediatric trauma: outcome prediction after severe injury. In revision
  • Timing of PICU deaths: deaths in working hours and off-hours, 2014 to 2026. Abstract

Acute kidney injury after cardiac surgery

Kidney injury after congenital heart surgery is frequent and under-recognized in neonates. We test whether models survive the move between centers and look for persistent-injury phenotypes.

  • Clinical records
  • Hemodynamics
  • Multicenter CS-AKI modelsForty gradient-boosting models across four US cardiac centers and 1,100 children. External performance varied widely between centers, a direct lesson on generalizability.PublishedPediatric Critical Care Medicine, 2026
  • PC4 NEPHRONEarly prediction and phenotypic stratification of persistent cardiac surgery-associated AKI in neonates, within the Pediatric Cardiac Critical Care Consortium.Under reviewCritical Care
  • AKI in the PC4 registry: machine learning risk prediction and phenotyping in the multicenter PC4 cohort. Ongoing
  • Persistent CS-AKI at Hacettepe: a single-center cohort testing persistence as a high-risk phenotype. Submitted
  • After the ICU: one-year kidney function and growth after CS-AKI. Protocol

Extracorporeal therapies and blood purification

Hacettepe started Türkiye's first pediatric ECMO and continuous hemodiafiltration programs. We measure what the circuit does to the child.

  • Extracorporeal
  • Dilution factor in pediatric CKRTCircuit-related hemodilution expressed as Vblood / (Vblood + VECBV), with an open bedside calculator on this site.PublishedPediatric Nephrology, 2026
  • Stem cells and AN69 membranesCD34-positive cells were found entrapped on a clogged CKRT membrane during allogeneic stem cell transplantation, yet engraftment was not compromised.PublishedArtificial Organs, 2025
  • Methotrexate toxicity and conventional hemodialysisThe first pediatric case series of high-dose methotrexate toxicity treated with conventional hemodialysis.AbstractPediatric Critical Care Medicine, 2022
  • Therapeutic plasma exchange in hyperinflammatory syndromes: systematic review and meta-analysis in children. In revision
  • Hacettepe ECMO cohort, 2013 to 2025: single-center outcomes reported with ELSO definitions. Ongoing
  • Methotrexate hemodialysis series: full manuscript. In revision

Language models for clinical text

Most clinical knowledge sits in free text, much of it in Turkish. We extract it with models that can run inside the hospital.

  • Clinical records
  • ICD-10 coding of Turkish PICU diagnoses: a language model translates and codes about 10,000 diagnosis strings. Ongoing
  • Electrophysiology reports: structured extraction from pediatric ablation and electrophysiology study PDFs. Ongoing
  • TurkishHealthBench: a Turkish adaptation of HealthBench for evaluating language models on health conversations. Ongoing

Trustworthy, private and reproducible clinical AI

Before a model reaches the bedside, we try to break it: adversarial inputs, leaked identity, and results that do not replicate.

  • Imaging
  • ECG
  • OCR-mediated modality dominanceText embedded in radiology images can steer vision-language model outputs, with implications for radiology AI trustworthiness; with MIT Critical Data.PreprintmedRxiv, 2026
  • Language models in critical decision-makingCommentary on balancing caution and innovation when large language models enter critical care.PublishedCritical Care, 2023
  • ECG identity leakage: whether patient identity survives in shared ECG representations, and whether removal strategies work. Preregistered. In preparation
  • Connectome-wired networks for chest radiographs: tests a fruit-fly connectome topology against rewired control graphs. In preparation
  • Hidden pixel injection in DICOM: a prototype for testing imaging AI pipelines against inputs invisible to readers. Prototype
  • Governance of ECG decision support: signal-aligned evaluation. Letter in preparation

ECG, echocardiography and sound

Low-cost signals that could screen children where pediatric cardiologists are scarce.

  • ECG
  • Imaging
  • Sound
  • PulmoVecTwo-stage stacking meta-learning on Google's HeAR foundation model for multi-task classification of pediatric respiratory sounds. Code.PreprintarXiv, 2026
  • EchoPLAX-SegU-Net segmentation of parasternal long-axis echocardiography views.CodeGitHub
  • PEEK newborn screening: digital stethoscope heart sounds and ECG to detect critical congenital heart disease in newborns, with Cleveland Clinic and Ankara Bilkent City Hospital. Ongoing
  • ECG-omics: 12-lead ECGs encoded as codon-like sequences to detect structural heart disease (EchoNext, MIMIC-IV). In preparation
  • PRIMED-AI: echocardiography and ECG foundation-model embeddings for ejection fraction when only ECG is available at deployment; with MIT Critical Data. Ongoing
  • Repaired tetralogy of Fallot: ECG age gap and unsupervised ECG phenotypes. In preparation

Emergency care, registries and population health

Forecasting demand and outcomes outside the ICU, from the emergency department to national registries.

  • Population data
  • Clinical records
  • SAFED-CTwenty forecasting models on 352,843 pediatric emergency admissions, MLOps retraining, and integer linear programming for physician shifts: about 4.3 fewer patients per physician at peak shifts without extra staff.PublishedJAMIA Open, 2025
  • Forecast-driven staffing in practiceProspective quasi-experimental pilot of the SAFED-C schedule in a pediatric emergency department.PreprintResearch Square, 2026
  • Escalation of care in childhood pneumoniaMachine learning prediction of escalation in hospitalized children, with a web app.In pressJMIRx Med
  • depremdoktor.comAn e-health platform for drug dosing and clinic coordination during the 2023 southern Türkiye earthquakes.PublishedDisaster Medicine and Public Health Preparedness, 2024
  • MATRA-AMulticenter study of massive transfusion.PublishedVox Sanguinis, 2021
  • Cystic fibrosis registry of Türkiye: one-year exacerbation and lung function prediction in modulator-naive children across 11 centers. Abstract, ECFS 2026
  • Hızır Graph: 113,051 pediatric ambulance transports in Ankara, 2020 to 2023, mapped with GIS. Manuscript complete
  • MicroPICU: antibiogram differences between care units and antimicrobial resistance forecasting. Ongoing
  • Economic crises and micronutrients: micronutrient status in children, 2005 to 2023. In preparation
  • Motorcycle crash mortality: national casualty records. In preparation
  • Medication preparation by PICU nurses: time-motion study. Abstract

Data infrastructure and standards

The groundwork that makes the other lines possible, in Ankara and in multicenter networks.

  • ECG
  • Hemodynamics
  • Waveforms
  • Ventilation
  • Clinical records
  • Extracorporeal
  • NAFISThe pediatric ICU integrated data and research platform described above.InfrastructureBAP 23145
  • Hacettepe PICU RegistryA web registry with an ECMO module aligned to ELSO definitions, CS-AKI and severity calculators; CKRT, plasma exchange and brain death modules in development.Livehpicu.app
  • Sickbay and PC4 pipelinesHigh-fidelity physiologic data retrieval and registry linkage at Cleveland Clinic, used in the methylene blue and low cardiac output studies.CompletedCleveland Clinic, 2024 to 2025
  • OMOP extension for pediatric CKRT: a common data model extension with an R package. In preparation
  • KafkaICU: mapping the hospital device telemetry stream (265 Kafka topics) ahead of NAFIS ingestion. Ongoing
  • Monitor capture on ESP32: reading GE CARESCAPE serial output with microcontrollers. Prototype
  • LSCData: a REDCap-style case report form platform for multicenter Turkish studies. In development

Bedside tools and apps

Small, open tools that answer a question at the bedside or in the classroom.

The complete list of articles, preprints, abstracts and book chapters, with abstracts, is on turkalpmd.com and Google Scholar.

How Claude will work in this lab

We are a small group of clinician-scientists building infrastructure that would normally need a software team. Claude extends that team. The boundary is fixed by our own protocol: researchers do not upload NAFIS data to external AI services, cloud language models included.

Hacettepe hospital network

Stays here

  • Patient-level physiologic and clinical data
  • Identity vault and encryption key shares
  • Model training and validation on real records
  • Queries, approvals and the audit log

What crosses

Reviewed before it moves

  • Source code, tests and infrastructure definitions
  • Schemas, data dictionaries and mapping tables
  • Synthetic and simulated signals
  • Aggregate tables and figures approved by the principal investigator

Claude

Works on these

  • Claude Code writes, reviews and tests our software
  • Claude drafts protocols, analysis plans and manuscripts
  • Shared projects keep standards, style and decisions consistent across the team

Never: identifiable records, waveforms or notes from real patients.

NAFIS engineering
The read-only Kafka consumer, idempotent ingestion, nightly Parquet compaction, catalog constraints, backup and restore verification, and the tests around each; container builds and version-controlled documentation, so the system outlives any one person's tenure.
Standards and interoperability
HL7 v2 (ADT, ORU, ORM) and FHIR R4 resources mapped to the OMOP common data model; LOINC, SNOMED CT, ICD-10 and ATC mappings; national Turkish health data dictionaries; our OMOP extension for CKRT.
Analysis code that can be audited
Modeling pipelines for kidney injury, low cardiac output and deterioration, written and reviewed against TRIPOD+AI and PROBAST+AI, then executed inside the hospital network.
Governance documents
Ethics protocols, data request and data use forms, consent and assent forms in age-appropriate Turkish, and variable dictionaries for each approved study.
Writing and review
Manuscripts, grant applications to TÜBİTAK and Horizon Europe, systematic review appraisal, and structured internal peer review before submission.
Teaching
Pediatric critical care fellows learn data science on simulated NAFIS data before they touch a real record.

We already build on Claude. The lab publishes open-source Claude skills for scientific writing and for agentic data analysis, and uses Claude Code to maintain the Life Support Center website and its bedside calculators, this page included.

Team and collaborators

Clinicians, engineers and methodologists in Ankara, Cleveland, Baltimore and Boston.

  • Izzet Turkalp Akbasli, MDLab director; NAFIS data architecture and integration lead
  • Prof. Benan Bayrakci, MDDirector, Life Support Center; head of the Division of Pediatric Critical Care; NAFIS principal investigator
  • Oğuzhan Serin, MDResearcher, Hacettepe University: digital health, clinical NLP, health technology assessment and AI governance
  • Orkun Baloglu, MDCollaborator, Cleveland Clinic: pediatric cardiac critical care and machine learning
  • Volkan Doğan, MScEngineering: computer vision engineer at Rapsodo
  • Prof. Jim Fackler, MDAdvisory board, Johns Hopkins University: clinical information systems and data-driven decision support in pediatric critical care

Frequent co-investigators include Selman Kesici and Özlem Tekşam at Hacettepe; Animesh Tandon, Bradley S. Marino and Samir Q. Latifi at Cleveland Clinic; Rohit S. Loomba; and Leo Anthony Celi at MIT Critical Data.

Funding, recognition and service

Funding and awards

  • Hacettepe University BAP infrastructure project 23145 for NAFIS hardware
  • Research support from Google Health during the Cleveland Clinic fellowship
  • Cleveland Clinic Catalyst award for a project on telemetric pediatric ICU data
  • Special Jury Award, Google Artificial Intelligence Competition (2022)
  • Amazon Web Services cloud and infrastructure award for depremdoktor.com (2023)
  • Contributor to NIH R01 and K award applications at Cleveland Clinic

Editorial and community

  • Associate editor, Journal of Pediatric Intensive Care
  • Reviewer for Critical Care Medicine, Pediatric Nephrology, BMC Medical Informatics and Decision Making, and Frontiers in Pediatrics
  • Monthly work with MIT Critical Data on trustworthiness and data governance in clinical AI
  • Invited speaker, American College of Cardiology Cross-Sectional AI Work Group (2025)
  • Mentor and panelist, Harvard Health Systems Innovation Lab Hackathon, Istanbul hub (2026)

Training

  • American Board of Artificial Intelligence in Medicine certification (2025)
  • Fundamentals of Implementation Science, University of Washington Department of Global Health (2025)
  • Good Clinical Practice (TÜSEB, 2026; Cleveland Clinic) and CITI human research ethics
  • MLOps and DevOps (Amazon Web Services); machine learning and generative AI (Databricks); data warehousing (Snowflake)

Contact

iakbasli@hacettepe.edu.tr

For data requests to NAFIS, the research board of the Life Support Center reviews proposals weekly; ethics approval is required before any access.

Digital Health and Artificial Intelligence in Critical Care Laboratory
Life Support Practice and Research Center
Hacettepe University İhsan Doğramacı Children's Hospital
Pediatric Intensive Care Unit
06230 Sıhhiye, Altındağ, Ankara, Türkiye
+90 312 305 13 50