Multi-Task Learning Mixed-Signal Classifier for

In-situ Detection of Atrial Fibrillation and Sepsis

Sudarsan Sadasivuni†1 , Sumukh Prashant Bhanushali†1 , Sai Srinivasa Singamsetti† , Imon Banerjee∗ ,
and Arindam Sanyal† (1 equal contributing first-authors)
∗ Departmentof Biomedical Informatics, Emory University, Atlanta GA 30322, USA.
† Department of Electrical Engineering, University at Buffalo, Buffalo, NY 14260, USA. Email: [email protected]

Abstract—This paper presents an on-chip analog machine

functions that leverage intrinsic analog nonlinearity, and 3) an
learning (ML) classifier IC for detecting atrial fibrillation (AFib) error-aware AI training methodology that trains the ANN with
and sepsis from electrocardiogram (ECG) signal. The proposed circuit models to ensure good match between software model
technique allows continuous in-situ health surveillance using and ANN circuit. Compared to SRAM based IMC, switched-
2021 IEEE Biomedical Circuits and Systems Conference (BioCAS) | 978-1-7281-7204-0/21/$31.00 ©2021 IEEE | DOI: 10.1109/BIOCAS49922.2021.9644994

wearables with embedded AI for early detection of underlying capacitor IMC has better linearity for VMMs (see Fig. 1(b)).
health issues. The analog classifier uses custom activation func-
tion and performs in-memory computation (IMC) with switched-
Linearity of VMM using SRAM cells is fundamentally limited
capacitor circuits for reduced data movement. Designed in 65nm, by nonlinear relationship between discharge current (Ids ) and
the test chip achieves average accuracy of 98.2% for AFib voltage on BL/BLB, and the VMM results are not linear over
detection, and 90.7% for predicting sepsis 4 hours before onset. the full dynamic range [1]. In addition, matching capacitors is
The energy efficiency of the test-chip is 12.9nJ/classification which easier than matching transistors and Ids across large SRAM
is 4× better than state-of-the-art.
array. The proposed classifier is demonstrated on single lead
Index Terms—Machine learning, atrial fibrillation, sepsis,
mixed-signal classifier and in-memory computation ECG data from Physionet 2017 dataset and on data from a
patient cohort admitted to Emory University Hospital (EUH)
between 2014 to 2018.
I. I NTRODUCTION
On-chip ANN total energy:
Atrial fibrillation (AFib) results in more than 150,000 un- (core area: 1.67mm2) 12.9nJ/classification

0.5mm
R peak 0.8mm
derlying cause of deaths in the USA annually. However, most ECG Signal
Layer2 Layer3
FE: includes
feature extractor
1.2mm

R
AFib patients are asymptomatic, leading to reduced awareness N-N
0.6mm 2.1mm
+ pre-processing
to remove
T P T
and less chances of managing stroke risks. Sepsis is another Q
Layer1 baseline wander
S
significant cause of death in the USA with close to 40% 2.5mm
Vth1 Vout>Vth1 Normal
mortality rate after onset and with 80% of the patients having DAC
Vth2<Vout<Vth1 AFib
onset outside hospital. Abnormalities in electrocardiogram Digital Feature
Extractor + DAC
Vout

(ECG) signal of patients can be used as an early indicator for pre-processing

(synthesized)
Output
neuron
Vth2 Vout<Vth2 Noisy

both AFib and sepsis onset. Continuous health surveillance DAC

Vout>Vth3 Sepsis
using wearables with built-in artificial intelligence (AI) is a Shared hidden layer 1 Shared hidden layer 2 Vth3 Non-sepsis
(20 neurons) (6 neurons) (a)
potential solution for risk management while securing patient SRAM-IMC with peripheral ADC Switched-cap IMC with analog activation
data privacy. However, AI analysis is typically computationally BL WL BLB
Vi[1] Vi[2] Vi[n]
intensive and it is difficult to embed AI model within resource ϕs ϕs ϕs ϕs ϕs ϕs

constrained wearables. Approaches to reduce energy con- Ids

C[1] C[2] C[n]
CS amp

sumption for AI analysis involve low-precision computation ∑ Vi [k] C[k] Vout

Vout = ϕs
∑ C[k]
and in-memory computation (IMC) to reduce data movement. ADC

State-of-the-art mixed-signal AI circuits have typically only 1. Ids is non-linear function of bitline voltage 1. Switched-cap MAC computation is highly linear
2. Random mismatch in Ids in each bitcell 2. Capacitors have better matching than transistors
demonstrated on-chip vector matrix multiplication (VMM) or 3. Peripheral ADC bottleneck 3. No ADCs in the ANN
4. ANN weights can be reprogrammed easily 4. ANN weights cannot be reprogrammed
the first hidden/convolutional layer, with rest of the AI model (b)
implemented in software [1], [2].
Fig. 1: a) Multi-task learning ANN for atrial fibrillation and
This work presents a mixed-signal, multi-task learning
sepsis prediction from ECG signal, b) comparison of IMC with
(MTL) classifier for detecting AFib and sepsis from temporal
SRAM vs switched-capacitor
ECG signal (Fig. 1(a)), with the 3-layer artificial neural
network (ANN) classifier implemented on-chip. The MTL
II. M ULTI - TASK LEARNING ANN T RAINING
ANN model gives high accuracy for both AFib and sepsis
prediction tasks, since both tasks are fundamentally identi- A. Description of dataset
fying abnormalities in ECG signal. The key contributions of The 2017 PhysioNet dataset comprises of ECG recordings
this work are - 1) demonstration of fully integrated analog lasting from 9 seconds to over 60 seconds. The ECG record-
MTL ANN with switched-capacitor IMC, 2) custom activation ings are sampled at 300Hz, and contain normal sinus rhythm,

978-1-7281-7204-0/21/$31.00 ©2021 IEEE

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on March 09,2023 at 08:05:00 UTC from IEEE Xplore. Restrictions apply.
AFib and noisy data of 5971 patients. The sepsis dataset is Vcm MAC
Xp[1] Xp[2] Xp[n]
obtained from EUH with approval from Emory Institutional (a)
ϕs ϕsb ϕs ϕsb ϕs ϕsb
Review Board,. The cohort consisted of 800 patients admitted output offset
to the ICUs at two hospitals within the Emory Healthcare ϕse
cancellation
C[1] C[2] C[n] Coff
system from 2014 to 2018. For each patient, there is at least Vcm
φs
Vom
Vcm Vop
8 hours of ECG signal recordings, sampled at 300Hz, from ϕse C[1] C[2] C[n] Coff custom tanh
the time of admission in the ICU. 400 patients in the cohort 1

had sepsis, with onset time assigned using Sepsis-3 criterion. ϕs ϕsb ϕs
0.5
ϕsb ϕs ϕsb

f(Vin) (V)
0

The goal of this work is to detect sepsis 4 hours before onset Xm[1] Xm[2] Xm[n] -0.5

-1
to allow adequate time for the 3-hour recommended sepsis Vcm -1 0 1
Vin (V)

treatments that have been shown to significantly improve

Vcm
sepsis outcomes [3]. Xp[1] Xp[2] Xp[n]
(b)
B. Feature extraction ϕs ϕsb ϕs ϕsb ϕs ϕsb

The input features are calculated on 30 second windows ϕse C[1] C[2] C[n]
Vcm
of the ECG signal, and only time-domain features are used Vo
Vcm
for low-cost hardware implementation. The time-domain fea- ϕse C[1] C[2] C[n]
custom softmax
tures are calculated from first-order statistical measures of
shape, dispersion, location and distribution of R peaks, QRS ϕs ϕsb ϕs ϕsb ϕs ϕsb

complexes, PR intervals, ST intervals, QT intervals, and NN Xm[1] Xm[2] Xm[n]

Vcm
intervals (see Fig. 1) in the single-lead ECG signal, and
involves calculation of standard deviation, mean, median and Fig. 2: Circuit schematic of a) hidden neuron with custom tanh
maximum/minimum values. 63 time-domain features are used activation b) output neuron with custom softmax neuron
for AFib detection, while a subset of 14 features, computed
on NN intervals and R peaks, are used for sepsis prediction. software co-design methodology [4]. The amplifier transfer
The feature extractor (FE) removes baseline wander from ECG curve and its derivative are imported into the ANN training in
signal by subtracting the median value from each segment. Matlab which starts with random weights. Stochastic gradient
C. ANN model training and circuit design descent function is used to optimize the ANN weights at each
epoch by minimizing the loss function. Once the ANN is fully
A three-layer MTL ANN is trained for detecting AFib and trained, the weights are encoded as capacitor values in the
predicting sepsis onset. The first two hidden layers have 20 MAC circuits. The ANN weights are quantized to 4-bit in
and 6 neurons respectively, and use a custom tanh activa- the hidden layers, and 6-bit in the output layer. The weight
tion function, while the output layer uses custom softmax quantization is done during the training iterations to preserve
activation. The result of softmax activation at the output accuracy during testing. 4fF unit capacitor, with mismatch
neuron is compared with threshold voltages for classification standard deviation of 0.4%, is selected as LSB weight in the
into ‘normal/AFib/noisy’ and ‘sepsis/non-sepsis’ categories MAC circuits to ensure that classification accuracy remains
as shown in Fig. 1. The threshold voltages for the two close to 99% even in presence of random mismatch.
prediction tasks are calculated during the ANN training phase
to optimize loss function for each task. Fig. 2 shows cir- III. M EASUREMENT R ESULTS
cuit schematic of neurons in the hidden and output layers. Fig. 1(a) shows the microphotograph of the test-chip with
Switched-capacitor circuits are used to perform in-memory, core area of 1.67mm2 . The FE, and DACs to convert digital
charge-domain multiply-and-accumulate (MAC) operations. features into analog signals, are implemented off-chip. The
Bottom-plate sampling technique is used to suppress charge on-chip ANN consumes 7.1µW for each inference while
injection. operating from 1.1V power supply at 1kHz, resulting in
The custom activation functions are realized using common- an energy consumption of 7.1nJ/inference. The DACs and
source differential amplifiers as shown in Fig. 2. The custom digitally synthesized FE are estimated to consume 3.8nJ and
tanh activation circuit uses fully differential common-source 2nJ respectively for each inference. Thus, the test-chip has an
amplifier with output offset cancellation as shown in Fig. 2(a). estimated energy consumption of 12.9nJ/inference. The energy
During sampling phase, φs , the differential inputs of the consumption will increase to 13.6nJ/inference if analog front-
amplifier are shorted together, and the offset is stored in the end amplifier and 14-bit ADC for digitizing ECG signal is
capacitor, Cof f , which is subtracted from the amplifier output integrated on-chip.
during evaluation phase. The custom softmax activation circuit
uses a common-source differential amplifier with single ended A. AFib detection results
output as shown in Fig. 2(b). The AFIb dataset is randomly split into 4767 training
To ensure that software training results with custom activa- samples and 1204 test samples. Fig. 3(a) shows the measured
tion functions match IC measurements, we apply a hardware- confusion matrix on the test set. The test-chip detects AFib

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on March 09,2023 at 08:05:00 UTC from IEEE Xplore. Restrictions apply.
with 98.8% accuracy, specificity of 1 and sensitivity of 0.89. TABLE I: Comparison with state-of-the-art ASICs
Threshold voltages for the class boundaries are calculated JSSC TBioCas JSSC ISSCC This
during foreground calibration step that applies the training 2019 [5] 2019 [6] 2020 [7] 2021 [8] work
samples to the test-chip and calculates the threshold voltages Process 65nm 180nm 40nm 65nm 65nm
to maximize classification accuracy on the training samples. Area (mm2 ) 5.9 0.92 0.24 1.74 1.67
Fig. 3(b) shows measured histogram of accuracy and sensitiv- Accuracy − 99.3% 96% 99.3% 98.2%1
ity for 1000 repeated evaluations on the test set. Small standard Type digital AMS2
deviation in accuracy and sensitivity for repeated evaluations Energy 0.33µJ 3.21µJ 51.6nJ 2.25µJ 12.9nJ
Model ANN ANN TDDL3 ANN ANN
demonstrate robustness against noise.
Class # 2 5 2 2 3
1
average of 4 chips, 2 AMS: analog/mixed-signal, 3
TDDL: task-
1 1023 driven dictionary learning
True Class

2 14 121

B. Sepsis prediction results

3 46

1 2 3
The sepsis dataset is split randomly into 620 training
Predicted Class samples and 180 test samples. Fig. 5 shows the measured
1: Normal; 2: AFib; 3: Noisy
accuracy and sensitivity of sepsis prediction as a function
Accuracy : 98.8%; Specificity: 1 of time before onset. As expected, the prediction accuracy
Sensitivity: 0.89
and sensitivity improves closer to onset. This work predicts
sepsis 4 hours before onset to allow sufficient time for sepsis
treatment [3].
Fig. 3: a) Measured confusion matrix for AFib dataset, b)
accuracy and sensitivity for 1000 evaluations 1
Accuracy
Sensitivity
Fig. 4(a) shows the measured accuracy and sensitivity 0.95
0.950.95
0.93

Acc/Sen
as the power supply voltage is swept from 1.2V to 0.8V. 0.92 0.92
0.90
0.91
0.9 0.90
Classification accuracy reduces with supply voltage. Fig. 4(b) 0.88 0.87
shows the measured accuracy and sensitivity for 4 test chips. 0.85 0.84 0.84
The class boundaries are calculated for each test-chip through
0.8
foreground calibration using training samples. The average 1 2 3 4 5 6
accuracy and sensitivity across 4 test-chips are 98.2% and 0.89 Time to onset (hours)

respectively. Table I compares our prototype with state-of-the- Fig. 5: Measured accuracy and sensitivity of sepsis prediction
art ASICs demonstrated on AFib detection tasks. The proposed before onset
ANN consumes the lowest energy thanks to analog ANN.
Fig. 6(a) shows the measured confusion matrix on the test
1.1 set. The test-chip predicts sepsis with 91.1% accuracy, speci-
Accuracy
Sensitivity ficity of 0.94 and sensitivity of 0.88. Similar to AFib dataset,
1 threshold voltage for the class boundaries are calculated on
Acc/Sen

0.98 0.98 0.99

0.96
the training samples to maximize prediction accuracy on the
0.90
0.9 0.88 0.88 0.89 0.89 0.90 train set. Fig. 6(b) shows measured histogram of accuracy and
sensitivity for 1000 repeated evaluations on the test set. Small
0.8 standard deviation in accuracy and sensitivity for repeated
0.8 0.9 1 1.1 1.2
Vdd (V) evaluations demonstrate robustness against noise.
(a)
1.1
Accuracy 1 85 5
True Class

Sensitivity

1 0.99
Acc/Sen

0.99 0.98
0.97
2 11 79

0.90
0.9 0.88 0.89 0.88 1 2
Predicted Class
1: Normal; 2: Sepsis
0.8
1 2 3 4 Accuracy : 91.1%; Specificity: 0.94
Chip # Sensitivity: 0.88

(b)

Fig. 4: Measured accuracy and sensitivity for a) as a function

Fig. 6: a) Measured confusion matrix for sepsis dataset, b)
of supply voltage, b) multiple chips
accuracy and sensitivity for 1000 evaluations

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on March 09,2023 at 08:05:00 UTC from IEEE Xplore. Restrictions apply.
 Fig. 7(a) shows the measured accuracy and sensitivity as the
power supply voltage is swept from 1.2V to 0.8V. Prediction
accuracy reduces with supply voltage. Fig. 7(b) shows the
measured accuracy and sensitivity for 4 test chips. The average
accuracy and sensitivity across 4 test-chips are 90.7% and 0.84
respectively. Table II compares our sepsis prediction work with
1
Accuracy
0.930.94
Sensitivity 0.90 0.91
0.9 0.88
0.86
Acc/Sen

0.82
0.80
0.8
0.73
0.7
0.64

0.6
0.8 0.9 1 1.1 1.2
Vdd (V)
(a)
1
Accuracy Fig. 8: Comparison with digital baseline and state-of-the-art
0.95 Sensitivity
0.92 0.91
Acc/Sen

0.90 0.90
0.9 0.88
distribute reprints for Government purposes notwithstanding
0.85 0.84
0.83
any copyright notation thereon. The views and conclusions
0.8
0.81 contained herein are those of the authors and should not
be interpreted as necessarily representing the official policies
0.75
1 2 3 4 or endorsements, either expressed or implied, of Air Force
Chip # Research Laboratory or the U.S. Government.
(b)
R EFERENCES
Fig. 7: Measured accuracy and sensitivity for a) as a function [1] Q. Dong et al., “ A 351TOPS/W and 372.4 GOPS compute-in-memory
of supply voltage, b) multiple chips SRAM macro in 7nm FinFET CMOS for machine-learning applica-
tions,” in ISSCC, 2020, pp. 242–244.
state-of-the-art. Our work has the highest prediction accuracy [2] H. Valavi, P. J. Ramadge, E. Nestler, and N. Verma, “A mixed-signal
4 hours before sepsis onset. To the best of our knowledge, binarized convolutional-neural-network accelerator integrating dense
weight storage and multiplication for reduced data movement,” in IEEE
there are no custom ASICs in the literature that perform sepsis Symposium on VLSI Circuits, 2018, pp. 141–142.
prediction. Fig. 8 compares our work with digital baseline [3] L. Pruinelli et al., “Delay within the 3-hour surviving sepsis campaign
synthesized in 65nm, and with state-of-the-art AI ASICs for guideline on mortality for patients with severe sepsis and septic shock,”
Critical care medicine, vol. 46, no. 4, p. 500, 2018.
different bio-medical applications. [4] S. T. Chandrasekaran et al., “Fully Integrated Analog Machine Learning
Classifier Using Custom Activation Function for Low Resolution Image
TABLE II: Comparison with state-of-the-art Classification,” IEEE TCAS–I, vol. 68, no. 3, pp. 1023–1033, 2021.
[5] S. Yin et al., “A 1.06-µ W Smart ECG Processor in 65-nm CMOS for
CinC CCM JAMIA Nature This Real-Time Biometric Authentication and Personal Cardiac Monitoring,”
2019 [9] 2019 [10] 2020 [11] 2021 [12] work IEEE JSSC, vol. 54, no. 8, pp. 2316–2326, 2019.
Accuracy 84.5% 67% − − 90.7%1 [6] Y. Zhao, Z. Shang, and Y. Lian, “A 13.34 µW Event-Driven Patient-
Sensitivity 0.66 0.85 0.84 0.86 0.841 Specific ANN Cardiac Arrhythmia Classifier for Wearable ECG Sen-
sors,” IEEE transactions on biomedical circuits and systems, vol. 14,
tonset 2 4 hours
no. 2, pp. 186–197, 2019.
Model LSTM 3 SM 4 RNN 5 RF6 ANN [7] K.-C. Chen, C.-Y. Chou, and A.-Y. Wu, “A Tri-Mode Compressed
1
average of 4 chips, 2 hours before sepsis onset; 3 LSTM: long-short Analytics Engine for Low-Power AF Detection With On-Demand EKG
term memory; 4 SM: survival model; 5 RNN: recurrent neural network; Reconstruction,” IEEE Journal of Solid-State Circuits, 2020.
6 [8] J. Liu et al., “BioAIP: A Reconfigurable Biomedical AI Processor
RF: random forest with Adaptive Learning for Versatile Intelligent Health Monitoring,” in
ISSCC, vol. 64, 2021, pp. 62–64.
IV. C ONCLUSION [9] B. Roussel, J. Behar, and J. Oster, “A Recurrent Neural Network for the
This work has presented an analog machine learning classi- Prediction of Vital Sign Evolution and Sepsis in ICU,” in Computing in
Cardiology (CinC). IEEE, 2019, pp. Page–1.
fier IC for AFib detection, and sepsis prediction from patient [10] S. Nemati et al., “An interpretable machine learning model for accurate
ECG signal. The combination of switched-capacitor IMC and prediction of sepsis in the ICU,” Critical care medicine, vol. 46, no. 4,
custom analog activation circuits results in 4× improvement p. 547, 2018.
[11] A. D. Bedoya et al., “Machine learning for early detection of sepsis: an
in energy efficiency without sacrificing prediction accuracy. internal and temporal validation study,” JAMIA open, vol. 3, no. 2, pp.
ACKNOWLEDGMENT 252–260, 2020.
This material is based on research sponsored by Air Force [12] K. H. Goh et al., “Artificial intelligence in sepsis early prediction and di-
agnosis using unstructured data in healthcare,” Nature communications,
Research Laboratory under agreement number FA8650-18-2- vol. 12, no. 1, pp. 1–10, 2021.
5402. The U.S. Government is authorized to reproduce and

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on March 09,2023 at 08:05:00 UTC from IEEE Xplore. Restrictions apply.