Jet engines are remarkable mechanical engineering feats and include more than 100, 000 parts. A rudimentary understanding of mechanical engineering concepts helped physicians such as William Harvey, who described the heart as a pump with 1-way valves, make fundamental advances in medicine in the 17th century. Now, well into the 21st century, jet engines are monitored far more intensively than people. This article explores how contemporary digital technologies, such as on-engine monitors or wearable sensors and artificial intelligence (AI), can be used to maintain the well-being of jet engines and humans.
Jet engines are built to very tight specifications and operate within well-defined bounds of normality. In contrast, with the human body, it is still not known what "normal" is for common metrics, such as blood glucose after eating or blood pressure throughout the day and night.¹ With jet engines, performance parameters (eg, temperatures, pressures, and shaft speeds) and vibration data are continuously monitored in flight to assess engine health. Advanced signal processing and machine learning are applied to the engine data to provide early warnings of incipient problems that would require unscheduled engine maintenance and take the plane out of service. End-of-flight summary reports that distill gigabytes of data are transmitted to the maintenance engineers on the ground.
Such intensive monitoring of people occurs today only in the hospital intensive care unit, with continuous monitoring of all vital signs and other parameters, such as oxygen saturation. Yet these data are processed with relatively primitive, rule-based algorithms that frequently trigger inappropriate alarms.
Wearable sensor technology has reached a stage of development such that the vital signs monitored in the intensive care unit can now be recorded unobtrusively and continuously by individuals who are healthy or not acutely ill. Wearable technology can also be used by patients with chronic diseases, such as diabetes, to track blood glucose levels with a calibrated sensor, obviating the need for fingerstick measurements. But algorithms for humans are mostly unidimensional; for instance, the glucose-centric monitors and algorithms in common use do not integrate food intake, physical activity, sleep, the gut microbiome, and other known meaningful inputs. Furthermore, the megabytes of patient data generated every 24 hours will only be a form of noise, unless the signal indicating a change in the condition of the individual or patient can be reliably extracted and is clinically meaningful. This requires approaches such as machine learning to generate a simple "person summary" similar to the flight summary from a jet engine.
Novelty detection is an approach that has been applied both to jet engines and patient monitoring. With novelty detection, machine learning algorithms are used to learn a model of normality from sensor data such that abnormal behavior for that engine or patient can subsequently be identified.² With airplane engines, novelty detection was first applied to the identification of abnormal vibrations during engine testing. In health care, the method has been used in the analysis of mammograms to identify cancerous masses and in critical care monitoring to provide early warning of physiological deterioration. With the latter, the novelty detection software was the first instance of machine learning approved by the US Food and Drug Administration for patient monitoring.³
A digital twin of a jet engine consists of a particular engine and a computer model that captures as accurately as possible the state of that engine. The engine and its model, usually the combination of a physics-based model and advanced analytics, are closely coupled via a range of sensors. The models, in combination with the sensor data from the actual engine, provide the ability to predict performance for different scenarios and, in turn, enable predictive maintenance to be carried out, with replacement of components before they fail.
When applied to people, the term digital twin can have 2 connotations. One, like the jet engine, is a computer simulation of an individual—in silico—that dynamically reflects the individual's molecular status, physiological status, and lifestyle over time. Rather than relying on a concept of the normal derived from population studies, deep learning techniques—ideal for processing a range of disparate data—are being developed to define normality using an individual's longitudinal data.⁴,⁵
An alternative concept of digital human twins relies on a massive information infrastructure that comprehensively characterizes each individual for demographics, biologic omics, physiology, anatomy, and environment, along with treatment and outcomes for medical conditions. Here the twin represents the nearest-neighbor patient, derived from algorithmic matching of the maximal proportion of data points using a subtype of AI known as nearest-neighbor analysis. The nearest neighbor is identified using AI analytics for approximating a facsimile, another human being as close as possible to an exact copy according to the patient's characteristics to help inform best treatment, outcomes, and even prevention. There could be more than 1 twin identified that closely match a person, especially if the data infrastructure is very large. While such a resource has not yet been developed, and was inconceivable until recently, it has the potential to create a true learning health system, with each individual having a digital twin(s) to help guide that person's health management.
Key Differences Between Jet Engine and Patient Health Monitoring
Health monitoring is at the core an inverse problem: by making measurements with sensors attached to the engine or human body, the aim is to infer the internal state of a highly complex safety-critical system. With jet engines, forward models (going from causes to effects using physics-based approaches) are well developed. It is possible to predict, using such a forward model, how internal changes, for example, changes in air flow or temperature at a particular point in the engine, will affect the external sensor data. This is not possible yet with patient health monitoring because of the lack of a comprehensive forward model for the human body. The Physiome Project is an attempt at such a model, using computational methods that incorporate biochemical, biophysical, and anatomical information on cells, tissues, and organs. Progress has been made but the aim to use the physiome models, with patient-specific parameters, as a quantitative statement of phenotype that can be incorporated into electronic health records has not yet been achieved.⁶,⁷
It is also important to recognize that health care data are intrinsically different from other types of data. Under the same operating conditions, a jet engine will respond in exactly the same way to control signals. With human beings, the mere fact of making a measurement can affect that measurement; for example, a substantial number of patients will have their blood pressure raised simply because a physician is making the measurement. A patient with chronic obstructive pulmonary disease being cared for in hospital for an exacerbation will have near-normal values of oxygen saturation because of the nebulizer or oxygen mask treatment. In many instances, the treatment will affect the parameters being measured by the sensors in complex, unpredictable ways, with variable delays. Furthermore, sensors can readily detect abnormalities in people without symptoms, such as atrial fibrillation, leaving uncertainty as to the value of such monitoring.⁸ The same holds true for detection in healthy people of pathogenic genomic variants that are not actionable.
As with jet engines, the full potential of health monitoring for people will only be realized when individualized models underpin the monitoring algorithms. If the sharing of anonymized health care data becomes the norm, the digital (human) twin identified through nearest-neighbor analysis may prove to be more effective at personalizing treatment and delivering improved outcomes than an in silico model of the patient, whether the latter is a data-driven model or a multiscale biochemical, biophysical, and anatomical model. No matter what model is developed and used, prospective validation that it promotes health, rather than exacerbates false alarms, will be vital.
1. Manrai AK, Patel CJ, Ioannidis JPA. In the era of precision medicine and big data, who is normal? JAMA. 2018;319(19):1981-1982.
2. Tarassenko L, Clifton DA, Bannister PR, King S, King D. Novelty detection. In: Boller C, Chang F-K, Fujino Y, eds. Encyclopedia of Structural Health Monitoring. Vol 2. Chichester, United Kingdom: Wiley; 2009:653-675.
3. Tarassenko L, Hann A, Young D. Integrated monitoring and analysis for early warning of patient deterioration. Br J Anaesth. 2006;97(1):64-68.
4. Hinton G. Deep learning—a technology with the potential to transform health care. JAMA. 2018;320(11):1101-1102. sref
5. Bruynseels K, Santoni de Sio F, van den Hoven J. Digital twins in health care: ethical implications of an emerging engineering paradigm. Front Genet. 2018;9:31-42.
6. Hunter P. The virtual physiological human: the physiome project aims to develop reproducible, multiscale models for clinical practice. IEEE Pulse. 2016;7(4):36-42.
7. Li X, Dunn J, Salins D, et al. Digital health: tracking physiomes and activity using wearable biosensors reveals useful health-related information. PLoS Biol. 2017;15(1):e2001402.
8. Steinhubl SR, Waalen J, Edwards AM, et al. Effect of a home-based wearable continuous ECG monitoring patch on detection of undiagnosed atrial fibrillation: the mstops randomized clinical trial. JAMA. 2018;320(2):146-155.
Article reference: Tarassenko L, Topol EJ. Monitoring Jet Engines and the Health of People. JAMA. Published online November 15, 2018.
Lionel Tarassenko, MA, DPhil
University of Oxford,
Oxford, United Kingdom.
Eric J. Topol, MD
La Jolla, California.
Published Online: November 15, 2018.
Conflict of Interest Disclosures: Dr Tarassenko reported receiving grants from Rolls-Royce PLC from 1996 to 2008 and personal fees from Sensyne Health and Oxehealth Ltd (of which he is a founder and stock owner). Dr Topol reported receiving personal fees from Dexcom, Verily, and Illumina.