from embedded dsp to embedded ai: making chronic patient monitoring intelligent and scalable
DESCRIPTION
From Embedded DSP to Embedded AI: making chronic patient monitoring intelligent and scalable. Naveen Verma [email protected]. Body Area Network Technology and Applications WPI. Healthcare: a problem of scale. NHE breakdown. Anticipated US GDP and NHE. - PowerPoint PPT PresentationTRANSCRIPT
Naveen [email protected]
From Embedded DSP to Embedded AI:making chronic patient monitoring
intelligent and scalable
Body Area Network Technology and ApplicationsWPI
2
Healthcare: a problem of scale
Aim is to translate scalability of technologyinto scalability of clinical responsiveness
Anticipated US GDP and NHE NHE breakdown
• 60% is due to hospital/clinical visits• 75% ($1.3T) is due to management of chronic disease and conditions (chronic disease causes 70% of deaths)
3
Clinical inference: what does this signal mean?
~7.5sec
Electrical onset of seizure
Convulsions
4
Why is it hard? (I)Signals from low-power chronic sensors are non-specific
[Suffczynski, Neuroscience’04]
Seizure onset
Single-photon emission computed tomography (SPECT)
EEG
5
Why is it hard? (II)
Background activity
[Gotman, EEG & ClinicalNeurophis.’81]
Patient A EEG:
Expression of disease states is variable over patients & time
SeizureOnset(patient specific waveform)
6
Constructing and applying high-order models
1) Efficient techniques for data analysis• Utilize methods from the domain of supervised machine-learning• Embed these in low-power platforms and devices
2) Availability of physiological data• Exploit patient databases from the healthcare domain• Exploit data acquisition capability of the sensor network
Enabling data-driven patient-monitoring networks through…
7
Towards devices: seizure detection IC
Feature Extraction Processor
Inst. Amp.
ADC
2.5mm
2.5m
m
Supply voltage 1VI-amp LNA power 3.5μWI-amp input impedance >700MΩI-amp noise (input referred, 0.5Hz-100Hz) 1.3μVrms
I-amp electrode offset tolerance >1VI-amp CMRR >60dBADC resolution 12bADC energy per conversion 250pJADC max. sampling rate 100kS/s
ADC INL/DNL 0.68/0.66LSB
ADC SNDR 65dBDigital energy per feature-vector 234nJFeature vector computation rate 0.5HzTotal energy/feature-vector per EEG channel 9μJ
[JSSC, April’10]
(Patient specific)
8
Impact of patient–specific modeling
8
0102030405060
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
0
2
4
6
1 3 5 7 9 11 13 15
00.20.40.60.8
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Late
ncy
(sec
)Fa
lse
Alar
ms/
Hour
Sens
itivi
ty
Patient Number
60
40
20
0
6
4
2
0
1
0.6
0.2
536 hrs of patient tests:Patient-specific learningNo learning [Wilson’04] Latency:
• Specific: 6.77 ± 3.0 sec• Non-sp.: 30.1 ±1 5 sec
False alarms:• Specific: 0.3 ± 0.7 /hr• Non-sp.: 2.0 ± 5.3 /hr
Sensitivity:• Specific: 0.93• Non-sp.: 0.66
[Shoeb, EMBC’07]
9
Computational energy: ECG beat detection
Classifier energy scales w/ model complexityreal patient data → complex models
[DAC’11]
10
Expanding the standards
www.continuaalliance.orgTelemetry
Patient data & modelsSelected
patient-specific data
Clinical experts(centralized &
limited resources)
11
Summary
Physiological expressions in low-power sensing modalitiesAre non-specific and variable from patient-to-patient
Data-driven methods provided systematic approaches for constructing high-order models
Need:
Algorithms(identify data instancesto present to experts)
Platforms(exploit algorithmic
structure for low energy & flexibility)
Networks(selective utilization of
clinical resources)