I’ll cover three areas on this wall: Diagnostic imaging aircraft co-pilots that organize the urgent imaging studies and perform an extremely reliable second look. Early-declining patients/sepsis detection technology that pulls vital signs, laboratory data and progress notes to alert early for treatment. Operational efficiencies/EMS dispatch analytics that allow for improved throughput and rapid response times from prehospital to inpatient settings. For each, I’ll describe how it works (computer vision, ML, NLP), the advantages, disadvantages, pictures and a short video of the technology.

Why does this matter to me? I’ve seen how mere minutes, and reliable information, can change outcomes. AI can augment limited clinical capacity and uniformity of quality, but only with validation locally, review for bias, accountability and strong privacy practices. I’m very enthusiastic about technology that truly augments care: triage that reveals the clue at the correct time, documentation that results in time being returned to the clinician while being cautious about over-elimination of trust in the technology, quality of the data. I hope to be able to distinguish what is clinically real from hype and learn how to use AI safely in moving towards patient care.

Machine vision models based on deep learning scan X-ray/CT/MRI studies to alert the clinician of likely abnormalities and rank the time-dependence of patients needing immediate attention, thus giving the radiologist a consistent second opinion. Technology: computer vision (CNN-based neural networks), supervised ML and GPU runtime analysis. Benefits: faster triage and consistent abnormality detection; fatigue relief due to high-volume worklists. Limitations: false positives and false negatives, inter-institution variability, and workflow issues in which reliance is upon humans. My perspective: terrific as a co-pilot when locally validated and integrated with standardized reporting, never to be used as an oracle.

Video : https://www.youtube.com/watch?v=dCDuMyzWS8Q

AI for Early Sepsis Detection

Machine-learning prediction models monitor laboratory parameters, vital signs, and clinical notes in the hospital and out-patient environment for identification of potential sepsis a few hours earlier than without ML prediction, thus providing an opportunity for earlier therapeutic antibiotic and fluids. Technology: refurbished models of gradient boosting or deep learning, with potential NLP for evaluation of free-text notes in patients. Benefits: earlier therapeutic initiation; lower potential mortality. Limitations: alert fatigue in healthcare workers, drift in calibration, known variations in generalizability, and transparency of models. My opinion: great potential for high impact if monitored properly post-deployment and tuned for local populations, otherwise noise.

Video:

https://www.youtube.com/watch?v=WVxmbZJ12s0

AI in EMS Dispatch & Operations

Predictive modeling and routing algorithms are applied to ambulance transport and placement to obtain optimal placement and transport for emergency service utilizing historical demand factors as well as the current present-day environment, to improve the analytics of hospital throughput. Technology: predictive modeling, operation optimization, geographic information systems and environmental input, with reinforcement learning sometimes used. Benefits: shorter response time, better coverage, less bottlenecking of inflow and outflow of patients from the ED. Limitations: data issues and concerns over data quality, fairness across neighborhood settings, CAD/EHR integration and training of healthcare workers. My position: great upside for access and also fairness initiatives if models are auditable and performance published.

Video:

https://www.youtube.com/watch?v=9vOGPuP-I3s

Such improvements mean nothing if we do not practice good ethics:

• Validation: Show that the devices work at each site and continuously check reliability after go-live.

• Bias & equity: Analyze how the devices perform by age, gender, language, race/ethnicity and neighborhood so that no group is under-treated.

• Transparency & accountability: AI offers options; humans decide. Make it plain how a recommendation was generated and who is accountable.

• Privacy & security: Safeguard personal health information and any ambient audio signals used for recording notes.

I welcome AI that reduces noise and provides back minutes in emergencies and routine care—speedier triage, improved handoff, less clicks. I fear black box alerts and “set-and-forget” applications. The future I envision is human-centered AI: simple use, measured in actual results (time-to-treatment, fewer misses), and continuously measured for drift and bias. If we can couple strong guardianship to useful devices AI can assist in providing safer more uniform care—without loss of the human bond that patients need most.

These societal effects are behind the way I will be using AI as a future clinician. I am excited where AI leads to returning time back to care teams, results in lowering the time to treatment, and allows for standardization of quality across zip codes. I am wary where the roll-outs are made without equity checks or as to the uses of privacy. The bar for me is simple: is it true that a certain technology, in a measurable way, improves patient-centered outcomes without widening the already mentioned disparities or eroding trust? If we take the good measurement of governance and apply it to the intelligent human-centered designs of the AI technologies, it would support the goal of efforts towards a desired better safety and more predictable quality of outcomes of medical care while maintaining the human connection which is the goal of medicine.