Healthcare and ambient intelligence

Shivajyoti Bhattacharjee, Vice President – Healthcare & Life Sciences, Cybage explains that Despite all the increasing advances in the healthcare space, a significant fact and observation are that medical errors are the third leading cause of death and highlights the ways to reduce the same

There is an effort going on with the academics and leading universities to illuminate the darkness of healthcare with ambient intelligence. During this, we are talking about how to advance healthcare in the following areas:

• Drugs and medicine
• Medical imaging
• Medical devices
• Genomics

When we reflect on healthcare trends over the last two decades, we can confidently say that our life expectancy has increased. Despite all the increasing advances in the healthcare space, a significant fact and observation are that medical errors are the third leading cause of death. The past twenty years’ data reflects critical pain points for clinicians – errors in unintended execution for healthcare. According to the US National Institute of Healthcare report, in 1999, 251,000 deaths/year from medical errors compared to 37,000 death/year from car accidents. It is a huge no. Why is this no. so huge? – We know its unintended injuries/fatalities.

Healthcare is a very complex process. It involves surgery rooms, ICUs, pharmacies, homes to pediatric, geriatric to neurology to many different areas and many scenarios in healthcare deliveries that errors may creep in. Despite multiple systems and rules in place, from surgical instrument errors to incorrect drug prescriptions, all types of errors can occur, yet humans are humans. Many spaces of healthcare are in darkness – complex human behavior in complex environments.

Self-driving cars – Combining modern sensors, algorithms, GPS, and modern technology, we have entered an era in which we use and enable the cognitive capability to have humans drive better or here drive by themselves. Let us get into an analogy between self-driving technology and healthcare delivery. The vision is to endow healthcare spaces with ambient intelligence via smart sensors and Machine Learning (ML) algorithms. The goal is to reduce the errors in healthcare delivery – in vital statistics measurements (BP, Set-up IV, Wound pressure, Stethoscope, respirator ON, Pump, Inject, Respirator OFF) to patients to OR.

• Transforming the physical space with sensing ability
• Recognition of all activities – ML treating the sensed data
• Integration of a complete clinical data ecosystem across physician text notes, radiology, dermatology, pathology, vitals, and genomics to build an AI-assisted hospital

There are two critical spaces to consider-hospital and daily living spaces. We have surgery rooms, intensive care units, and other clinical spaces in hospital settings. Measuring patient mobility in the ICU using a simple sensor can be useful. Fine-grained healthcare activities pose technical difficulties. Hospital-acquired infection related to lack of Hand hygiene is a leading cause of patient fatality in US hospitals. More than 90,000 patients die every year of hospital-acquired infections leading to hand hygiene importance. Similar to mobility detection, it isn’t easy to implement hand hygiene. RFID is a potential solution, but it’s coarse and doesn’t tell if clinicians use hand hygiene. Depth sensors put above alcohol-gel dispensers could be one answer. Before and after leaving patient rooms, clinicians must wash their hands. We can utilise a Convolutional Neural Network (CNN) for activity detection. Algorithms can work on real-world hospital data and provide continuous monitoring data. Build models with methods to detect and predict these movements of clinicians’ touchpoints.

The need of the hour is to build ambient intelligence in activity recognition in defining the WHAT. Clinicians and family members can further interpret descriptive analytics of clinical status.

• Infection – fever, urinary frequency, respiratory rate
• Mobility – Falls, slowed movements, unstable transfers, front door loitering, immobility
• Sleep – sleeping, day/night reversal
• Diet – eating, fluid intake, alcohol consumption, pill consumption

Fine-grained healthcare activities are challenging technical problems. A lot of work has already happened in activity recognition:

• Action classification (What)
• Temporal action detection (What and When)
• Human object interaction (What and How)

Complex activities often involve multiple humans utilising a variety of objects leading to scenarios like –

• Multiple objects
• Multiple actions
• Multiple actors
• Multiple steps

Leading research in activity recognition has led to MOMA – A new framework for multi-object multi-actor (MOMA) activity parsing. A novel architecture for activity parsing using HyperGraph Convolutional Networks (HGCN). The following steps are involved in HGCN in sequence:

• Graph generation
• Graph encoding
• Instance aggregation modules
• Atomic action

For role classification, the video encoder creates a temporal module consisting of :

• Activity classification
• Sub-activity classification
• Atomic action

We need to be cognizant of technology’s ethical, privacy, and social implications even with the best intentions with implications on privacy and security with stakeholders. Technology is only a part of the solution. We need to keep in mind as clinicians, patients, and society need to come together to complete the solution loop in healthcare delivery. Some of the ways to protect privacy would be:

• Facial blurring in data scrubbing for visual data just like PII data masking
• Human body masking converts people to 3D modules to mitigate personal information and identity.
• Differential privacy adds noise to individual records.
• Using federated learning not to centralise the data by having data training on the edge instead of having trained at the centralised cloud server
• Having homographic encryption-stronger privacy guarantees across hospital servers and cloud computing providers.

For data and algorithmic fairness issues, we have tools for measuring algorithmic fairness, such as:

• Equal parity
• Proportional parity
• False-positive parity – Desirable when your interventions are punitive
• False-negative parity – Desirable when your interventions are assistive/preventive

The following would be recommended steps to implement:

• Step-1: IRB approval/minimise population bias/proxy consent for cognitively impaired results and recruit participants
• Step-2: HIPAA compliant platform in data collection and storage with data and disk encryption
• Step-3: Face and private area blurring. Annotate the data.
• Step-4: Edge computing/federated learning. Develop and deploy the model

A multistakeholder approach:

AI technology must collaborate with clinicians, patients, policy leaders, ethicists, and legal experts to answer all questions, mitigate risks and provide end-to-end optimal and effective new generation healthcare delivery.