Healthcare Use Cases

ECG Arrhythmia Detection & Classification (Heart Monitoring)

ECG monitors the electrical rhythms that drive heart function and is the key tool in cardiology for assessing whether the heart is in a normal rhythm state or one of numerous abnormal rhythms each with associated known conditions. ECG can be single or multi-lead each producing a time series typically sampled at ~200 Hz. 

Deep learning can be used to train a model that reads the time series ECG and assigns one of as many as 14 classes of rhythm type to each ~1 second time interval. The model is trained with ground truth labelled by cardiologists or certified cardiology technicians - and recent results based on large amounts of training data have shown the deep learning model's performance exceeds the accuracy of the average of a group of board certified MD cardiologists. (see recent Nature paper reference to right).

Medical Imaging - Disease Detection

3D imaging techniques such as MRI & CT are most often used tools for detection of disease or injury in internal organs or tissue. 2D or 3D Convolutional Neural Networks can be trained to precisely detect abnormal regions in medical images with an accuracy approaching or in some cases exceeding the average human trained physician.

Detection of Diabetic Retinopathy

Diabetic retinopathy and diabetic macular edemais are the leading causes of blindness in working age adults. Convolutional neural networks have proved capable of providing automatic detection and grading of these conditions that can help ophthalmologists in both decision support and in design of effective treatment plans. (See academic papers on this topic to right).

EEG diagnostic classification, & Brain-Computer Interface (BCI)

EEG uses a system of externally applied leads to monitor electrical brain activity. A prevalent use is to monitor for epileptic seizure and pre-seizure activity. A large clinical dataset (Temple University Abnormal EEG Corpus) of labelled data is available for training machine learning models. Deep learning methods have been developed specifically for this use case. 

EEG is an underutilized clinical technique with great potential for use in diagnostics & precision medicine  application to other illnesses such as depression and other psychiatric illnesses. A key limitation is the difficult and time consuming process of interpreting EEG signals which can have 21 or more concurrent signals at ~200 HZ.  Deep learning models have the potential to open up EEG's use to new clinical applications.

 

BCI (Brain-Computer Interface) is an emerging application where EEG signals are literally used to understand the human subject's thoughts (at least covering a limited scope of topics)  - the chief application being to assist and understand the needs of incapacitated patients. Other uses could be to act as a command interface between a person and machine in a way that is more effective than keyboard or speech recognition. 

Hospital Stay Predictive Analytics

Machine learning has been very successfully used to predict outcomes for hospitalized patients, such as: probability of re-admission, risk of sepsis, and risk of other conditions all based on standard available input data in their hospital records. Clearly this is applicable to improving care while reducing healthcare costs overall

Diagnostic Predictive Analytics based on Electronic Health Records (EHR)

Electronic Health Records represent the ultimate technical challenge in 'unstructured data': these records contain a huge amount of heterogeneous data sources that do not fit into a simple tabular database paradigm. EHR include physician notes and summary diagnoses, diverse test results, including imaging results (MRI, CT, and X-ray), recorded time series (ECG, vitals, ...), Genetic test results, and patient illness history and demographic data  - and all of this over time periods spanning potentially decades. 

So, one grand challenge in the application of data science to healthcare is to use these complex unstructured EHR as input data into a master predictive model, to train the model using huge amounts of historical data, and then predict future patient likely health status and it's time progression for a large number of illnesses. This diagnostic capability can be used as "decision support" for physicians, but also for proactive / preventive care, and can be used in communities or places where immediate physician support is scarce or non-existent.

Precision Medicine:  The Grand Challenge & Path to More Effective Treatment

Many of the use cases above use data science to predict the presence of an illness or disease state. The next step is to create models which can prescribe the most effective course of treatment among many available drugs or different treatment regimes. 

 

How does this work? Similar to our description of industrial "machine health" and predictive maintenance (see for example, I-IOT uses cases) the trained machine learning models for predictive diagnostic analytics produce an internal "health state vector" which is used to classify the probability of an illness. However, this health state vector additionally contains very detailed information on the similarity of a given patient to other patients seen in the model training. If historical data is available on the response efficacy of treatment from those patients, then a new model can be trained to recommend the most likely effective treatment. Similar, to consumer recommendation engines which find latent similarities between consumers and use them to predict similar responses to products - this applies the same rationale to treatment response. This hypothesis of using similarity between health state vectors to uncover patient population cohorts that will respond similarly to treatment courses must be proven in any given case, but these relationships are very strong and often allow building powerful and useful models.