Using machine learning to predict deterioration of patients, in this case paediatric patients, is something that will likely be the norm in the future landscape of healthcare.This study looks at machine learning as a tool to predict decompensation of paediatric patients in the ward thus needing ICU stay.
Technology has always been touted to ‘mend’ the failures of Man. Medical errors are common and often under-reported and unknown. Patients can sometimes pay the ultimate price for such failures in the healthcare system. The question remains, ” Can technology be depended upon to ensure patient safety?”
Let’s be honest, computers are a product of Man and often times it depends on correct input of data, in order for it to be accurate. However, the ability to process large amounts of data and analyse them in a short period of time, gives computers an advantage and the almost supernatural power of ‘seeing’ things that otherwise the human brain cannot comprehend or predict. Unfortunately, if it is being fed incorrect information, then there is little use in the output that computers will provide. Creating a new technology often will create new, previously unknown errors.
This is where artificial intelligence and computer learning comes into play. Similar to a child that learns from trials in life, computers can have the ability to learn and perhaps do it better and faster than Man can.
Humans are a very complex species. The way a brain processes information is complicated with a mixture of interpretation of its senses, past experiences, its view of the world and the array of emotions, creating a ‘potion’ that can be very difficult to replicate.
So what does the future hold? Are computers with artificial intelligence going to morph and take over the world? Well, likely not in the near future.
In healthcare, although information is important, the interaction with another empathetic human being is sometimes the only therapy that is desired. The non-verbal gesture is a power that cannot be underestimated in that doctor-patient relationship, which should still be considered the core of the healthcare industry. This is the moment where the patient has contact with the healthcare world and where some of the most private of information is shared.
There is no denying that digital technology powered by machine learning and artificial intelligence will soon power many of the functions of healthcare to boost efficiency and productivity while minimising errors and wastage. However, the true digital advocate will still remind everyone that the power of a human touch should never be downplayed.
EARLY PREDICTION OF PATIENT DETERIORATION USING MACHINE LEARNING TECHNIQUES WITH TIME SERIES DATA
Sareen Shah, David Ledbetter, Melissa Aczon, Alysia Flynn, Sarah Rubin
Learning Objectives: ere are currently no dynamic methods of identifying children on the pediatric ward at risk for decompensation and requiring intensive care unit (ICU) admission. We have developed a new early warning tool using machine learning techniques that is able to analyze time-series data. is tool enables calculation of trajectory of risk over time with dynamic data (continu- ously changing values) instead of relying solely on static data (values at one point in time). Methods: We conducted a retrospective cohort study of patients admit- ted to a tertiary care pediatric academic center between 1/2011 and 12/2015. One-hundred thirty-seven variables including vital signs, laboratory data, and other clinical parameters were extracted. A recurrent neural network algorithm was used to generate a predictive model for patients requiring transfer to the ICU within the next six hours. e model parameters were trained using 75% of the cohort, and the parameters were then tested on the remaining 25% of the cohort. We compared our model’s performance to our hospital’s Pediatric Early Warn- ing Score (PEWS) using the Area Under the Receiver Operating Curve (AUC). Results: Data from 43,279 patient encounters were used to generate the model. e model was tested on 14,719 encounters, of which 402 were transferred from the inpatient ward to the ICU. e recurrent neural network model yielded an AUC of 0.800 [0.773 – 0.826 CI] for patient transfer to the ICU within the next six hours, while the PEWS model had an AUC of 0.762 [0.734 – 0.790 CI]
(p = 0.026). Conclusions: Our machine learning model was more discriminating than PEWS scores for patients transferring from the pediatric ward to the ICU within six hours.