Home Search Collections Journals About Contact us My IOPscience

Kernel density estimation-based real-time prediction for respiratory motion

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2010 Phys. Med. Biol. 55 1311

(http://iopscience.iop.org/0031-9155/55/5/004)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 146.201.208.22
This content was downloaded on 10/09/2015 at 18:40

Please note that terms and conditions apply.

IOP PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 55 (2010) 1311–1326 doi:10.1088/0031-9155/55/5/004

Kernel density estimation-based real-time prediction

for respiratory motion

Dan Ruan
Department of Radiation Oncology, Stanford University, Stanford, CA, USA

E-mail: [email protected]

Received 16 April 2009, in final form 15 November 2009

Published 4 February 2010
Online at stacks.iop.org/PMB/55/1311

Abstract
Effective delivery of adaptive radiotherapy requires locating the target with
high precision in real time. System latency caused by data acquisition,
streaming, processing and delivery control necessitates prediction. Prediction
is particularly challenging for highly mobile targets such as thoracic and
abdominal tumors undergoing respiration-induced motion. The complexity
of the respiratory motion makes it difficult to build and justify explicit models.
In this study, we honor the intrinsic uncertainties in respiratory motion and
propose a statistical treatment of the prediction problem. Instead of asking
for a deterministic covariate–response map and a unique estimate value for
future target position, we aim to obtain a distribution of the future target
position (response variable) conditioned on the observed historical sample
values (covariate variable). The key idea is to estimate the joint probability
distribution (pdf) of the covariate and response variables using an efficient
kernel density estimation method. Then, the problem of identifying the
distribution of the future target position reduces to identifying the section in the
joint pdf based on the observed covariate. Subsequently, estimators are derived
based on this estimated conditional distribution. This probabilistic perspective
has some distinctive advantages over existing deterministic schemes: (1) it
is compatible with potentially inconsistent training samples, i.e., when close
covariate variables correspond to dramatically different response values; (2)
it is not restricted by any prior structural assumption on the map between the
covariate and the response; (3) the two-stage setup allows much freedom in
choosing statistical estimates and provides a full nonparametric description
of the uncertainty for the resulting estimate. We evaluated the prediction
performance on ten patient RPM traces, using the root mean squared difference
between the prediction and the observed value normalized by the standard
deviation of the observed data as the error metric. Furthermore, we compared
the proposed method with two benchmark methods: most recent sample and
an adaptive linear filter. The kernel density estimation-based prediction results
demonstrate universally significant improvement over the alternatives and are
0031-9155/10/051311+16$30.00 © 2010 Institute of Physics and Engineering in Medicine Printed in the UK 1311
1312 D Ruan

especially valuable for long lookahead time, when the alternative methods fail
to produce useful predictions.
(Some figures in this article are in colour only in the electronic version)

1. Introduction

Modern radiotherapy systems are capable of delivering the prescribed dose to a specified
target position with high precision. However, target motion, if not properly accounted
for, compromises the delivery accuracy. Intrafractional target motion during treatment is
managed with passive gating (Keall et al 2002) or adaptive tracking (Nuyttens et al 2006).
Furthermore, intrinsic system latency exists due to hardware limitation, software processing
time and data communication. These considerations have motivated a large body of studies on
prediction algorithms during treatment, especially for highly mobile targets, such as thoracic
and abdominal tumors that are affected by respiratory motion.
The problem of predicting respiratory motion has been intensively studied, and the existing
methods can be classified into two categories: (1) those that assumes a specific inference model
structure between the covariate, which is usually constructed from a finite sequence of historical
samples, and the response (2) those that are ‘model free’. The general class of linear filters
fall into the first category, even though they may differ in adaptivity and specific formulations.
Autoregressive moving average (ARMA) model is a straightforward generalization of the
linear model (McCall and Jeraj 2007). Linear models in the statistical setting have given rise
to the application of the Kalman filter and multiple models (Putra et al 2008, Zimmerman
et al 2008, McMahon et al 2007). By estimating the regression coefficients, these models
extrapolate the behavior between the training covariate and training response values, and apply
the estimated inference map to the test covariate.
One potential drawback of these fixed-structure models is their incapability to learn
the inference pattern from samples that are further away in space/time, but more similar
in behavior, e.g., training samples that are obtained at similar breathing phase. Changing
the representation space, such as the sinusoidal (Vedam et al 2004) or scale space (Ernst
et al 2007), is one way to ‘pull’ these samples together. An even more flexible way is to
learn the covariate–response directly, as in neural networks (Isaksson et al 2005, Kakar et al
2005, Murphy and Dieterich 2006). Note that even in neural network setup, a consistent and
smooth inference map is assumed, and it will have difficulty if there exist multiple inconsistent
training samples whose covariates are identical (or very close), but have very distinct response
values.
In this study, we adopt a statistical prospective and consider the probability distribution of
the test response upon observing the test covariate. Treating the prediction as a random variable
not only honors the inconsistent training samples, but also provides a natural nonparametric
description of the uncertainty associated with any prediction estimate. We adopt a kernel
density estimator to approximate the joint probability distribution of the covariate and
response variable from training samples. The distribution of the test response variable is
obtained as the section of the joint distribution picked out by the observed test covariate
value.
We present the basic theory of the proposed methodology in section 2. Section 3 describes
the test data and the implementation procedure and reports the experimental results. Section 4
provides some structural discussion, and section 5 summarizes the study and discusses future
work.
Kernel density estimation-based real-time prediction for respiratory motion 1313

2. Methods

2.1. Basic setup

At current time instant t, we are given a set of discrete samples si, i = 1, 2, . . . , K, of
breathing trajectory, acquired at preceding times ti < t prior to t. For simplicity, we discuss
the formulation when scalar observations are acquired at uniform time intervals and the look-
ahead length is an integer multiple L of the sampling interval. Then, for any i  K − L, we
can construct a length p covariate xi = [si−(p−1) , si−(p−2) , . . . , si ] and response y i = si+L .
The parameter  is an integer that indicates the ‘lag length’ used to augment the covariate.
It should be chosen properly to balance the effect of system dynamics and observation noise
(Ruan et al 2007).
Let z i = (xi , y i ) ∈ p+1 , i = 1, 2, . . . , M, denote the collection of training samples,
where xi and y i are the covariate variable and the response variable, respectively. The goal
of prediction is to obtain an estimate of the unknown y given a test covariate x.
The key idea of the proposed statistical method is to consider the probability distribution
(pdf) of the random vector Z = (X, Y ) ∈ n , and regard each training sample z i as
an realization of Z. The prediction (or more generally inference task) becomes two folds:
(1) estimate the conditional distribution p(y|x), for observed covariate x and (2) obtain an
estimate ŷ from such distribution. Rather than formulating the whole problem in a single
optimization setting, we have chosen to present the estimation of conditional pdf and the
subsequent estimation of y in a decoupled manner, to honor the fact that each module is
self-contained and that once the conditional distribution p(y|x) is obtained, the user has the
freedom to choose any estimate based on the specific application. Yet the overall setting of
constructing the distribution in the joint (X, Y ) space, and the logic of obtaining an estimate
based on the conditional probability holds regardless.
We present the framework for kernel-based prediction in section 2.2, which provides a
pdf of the response variable. Section 2.3 discusses some natural estimates derived from the
resulting distribution. Section 2.4 provides a specific example for the proposed scheme and
section 2.5 explains in depth certain design considerations, and describes some variations that
could potentially improve the prediction performance.

2.2. Estimating the distribution of response variable with kernel density approximation
We consider the samples z i = (xi , y i ), i = 1, 2 . . . , M, as independent samples of the
random vector Z and obtain the kernel density approximation (Duda et al 2001) of the pdf
p(z) as the superposition of (equally weighted) local kernel densities centered about each
sample z i :
1 
M
p(z) = κ(z|z i ), (1)
M i=1
where κ(z|z i ) is a local density kernel.
The distribution of the response variable conditioned on the covariate p(y|x) can be
written as
p(y|x) = p((x, y)|x) = p(z|x) = p(z)/p(x), (2)

where p(x) is the marginal distribution p(x) = p((x, y)) dy. The conditional distribution
p(y|x) is the (normalized) section of p(z) at X = x.
Equations (1) and (2) provide the principle for estimating the distribution of the response
variable conditioned on the observed test covariate, by approximating the joint distribution
1314 D Ruan

0.8

covariate
0.6

0.4

0.2

response

(a) single kernel κ(zz| z i ) (b) discrete training samples z i

(c) kernel density estimated joint distribution p(zz) (d) conditional distribution p(y | x).

Figure 1. Schematic of the kernel density estimation-based prediction method: a single kernel
density (a) is placed at each discrete training sample (b) to construct the joint pdf (c). Based on
the observed test covariate, the corresponding section of the joint pdf is selected and renormalized
to obtain the conditional distribution of the test response (d), where certain statistics could be used
as the prediction, such as mean, medium and mode, as illustrated in (d).

of covariate–response with kernel densities. Figure 1 illustrates the major component in this
procedure.
A common choice of kernel density is the Gaussian kernel:

1
κ(z|z i , i ) = N (z − z i ), i = exp[−(z − z i )T i−1 (z − z i )].
(2π )n/2 |i |1/2

In the prediction application, the various coordinates in z correspond to different states in the
covariate or response variable. Therefore, it is feasible to assume the covariance matrix to be
block separable:
 
x,i 0
i = 2 . (3)
0 σy,i

The covariance of the local Gaussian kernel about each sample z i reflects contribution of the
sample to the local curvature of the overall probability distribution. If all training samples are
obtained under similar environment, then it is reasonable to assume x,i = x and σy,i = σy
for all i. The block diagonal form of the covariance indicates a separable Gaussian kernel,
which enables further simplification of (2) as follows:
Kernel density estimation-based real-time prediction for respiratory motion 1315

1 
p(y|x) = κ((x, y)|z i )
M i
1   
= exp −(x − xi )T x−1 (x − xi ) − y − y i /σy2
C i
1   
= exp −(x − xi )T x−1 (x − xi )] exp[−y − y i /σy2 . (4)
C i

The normalization parameters C is independent of both i and y, so one could delay the
normalization until the last step by scaling the final conditional pdf to unity integral. Define
the sample weights wi as
 
wi = exp −(x − xi )T x−1 (x − xi ) , (5)

and (4) reduces to

1   
p(y|x) = wi exp −y − y i /σy2 , (6)
C̃ i

with normalization parameter C̃. This indicates that conditioned on the test covariate x, the
distribution of the response variable y is a Gaussian mixture, with each Gaussian component
centered at the sampled y i ’s. The weights in the mixture are determined by how ‘close’ the
test covariate x is to the covariate xi in the training samples.

2.3. Natural estimates

Given the distribution of the response variable Y conditioned on the observed covariate state
x, one could devise estimates of the response variable accordingly.

2.3.1. The mean estimate. The mean of a random variable minimizes the expected squared
error:

ŷ mean = arg min E[(y − Y )2 |x] = E[Y |x]. (7)

y

Given the Gaussian mixture conditional distribution (6), the mean estimate reads
 
1   1 
ŷ mean = E[Y |x] = ywi exp −y − y i  σy2 dy = wi y i . (8)
C̃ i i wi i

It turns out that the mean estimate is the weighted sum of the training response values, where
the weights are determined by the ‘closeness’ between the testing and training covariates.
Also note that explicit normalization is no longer necessary due to cancellation.

2.3.2. The mode estimate. The mode can be considered as a maximum a posteriori
probability (MAP) estimate, as it seeks a response value that maximizes the conditional
distribution:

ŷ mode = arg max p(y|x). (9)

y
1316 D Ruan

2.3.3. The medium estimate. From the point of view of order statistics, one could also adopt
the medium estimate, which is defined as
ŷ medium = {y : p(Y  y) = 1/2} . (10)
Note that for mixture Gaussian distribution in (6), the cumulative distribution can be obtained
as
 y 
 
p(Y  y|x) = ỹwi exp −ỹ − yi  σy2 dỹ
−∞ i

1 y − yi
= wi 1 + erf √ , (11)
2 2σy
where the Gauss error function is defined as
 x
2
e−x dx.
2
erf(x) = √
π 0

2.4. An exemplary scheme for respiratory motion prediction

As an example, we provide the algorithmic flow chart using Gaussian kernel and the mean
estimate.

Algorithm 1 Predict ŷ from (xi , y i ) with Gaussian kernel and mean estimate.

1: Determine covariance x and σy for covariate and response variables.

2: Compute the weighting according to (5).
3: Compute the mean estimate ŷmean = wi ywi i , equivalent to (8).
i

Note that it is unnecessary to compute the conditional distribution explicitly when the
mean estimate is used, as the order of taking the expectation to obtain the mean and computing
the weighted sum for the conditional pdf can be interchanged. This is not true in general, as
in the case of mode and medium estimates.

2.5. Design considerations and potential variations

We describe some design considerations and variations that could potentially improve the
prediction performance.

2.5.1. Data-driven covariance/bandwidth selection. In introducing the principle of the

kernel method, we assumed the covariance of the covariate and the response variables a priori.
In practice, one needs to determine these values. Setting the covariance for a general kernel is
a challenging problem and can be identified with the bandwidth selection problem studied in
nonparametric density estimators. To this end, statistical methods in data-driven bandwidth
selection (Sheather and Jones 1991, Botev 2006) may be applied.
In our application, the end goal is to obtain an estimate of the test response rather than
estimating the pdf itself and we expect the prediction performance to be less sensitive to the
choice of covariance than in general fitting problem. With the assumed separable Gaussian
kernel and approximate spatial invariance (cf (3)), we estimate the covariate covariance x
and response σy from the training population as follows:
Kernel density estimation-based real-time prediction for respiratory motion 1317

 1   1 
x̄ = xi , ȳ = yi ;
Ntrain i∈ training set
Ntrain i∈ training set
1  1 
x = (xi − x̄)(xi − x̄) , σy2 = (y i − ȳ)2 .
Ntrain − 1 i∈ training set Ntrain − 1 i∈ training set
(12)

2.5.2. Inhomogeneous Gaussian kernel with varying covariance. When the training samples
differ in their uncertainty, which may be caused by varying local noise level, one could assign
different covariances to different samples. In particular, samples with higher uncertainty
should use a kernel with higher covariance, corresponding to a ‘more flat’ local kernel. With
i being the covariance for the local kernel around sample z i , the conditional distribution in
(4) becomes
1 
p(y|x) = κ((x, y)|z i )
M i
1  1  −1

= exp −(x − xi )T x,i (x − xi )
M i (2π ) |x,i | σy,i
n/2 1/2
 
× exp −y − y i  σy2 . (13)
We modify the definition of sample weights wi as
1  −1

wi = exp −(x − xi )T x,i (x − xi ) , (14)
|x,i |1/2 σy,i
and the conditional distribution p(y|x) maintains the Gaussian mixture form as in (6).
Subsequently, the mean estimate ŷ mean is again a weighted sum of the sample response y i . The
only difference incurred by allowing the local Gaussian kernels to have different covariances is
that the local confidence level modifies the weights: less reliable training samples (with higher
|x,i |1/2 σy,i ) receive less weight. In data-driven approaches, neighborhoods with few samples
have higher uncertainty, and it is desirable to ‘flatten’ out the local kernel. This approach is
similar in spirit to the variable kernel method in Silverman (1986).

2.5.3. Robustness to outliers in training samples. Training samples during abrupt (and non-
repetitive) changes, such as patient coughing, are less representative of the general covariate–
response behavior and may be regarded as outliers. To increase the robustness of the kernel
density estimator to such outliers, one should decrease the contribution of a training sample
if its behavior differs significantly from other samples. We follow a similar philosophy of
iterative weight assignment for robust local weighted regression (Cleveland 1979, Ruan et al
2007) and adjust the weight of a sample in the kernel density estimation as follows.
Let B(·) be a scalar function that satisfies
• B(x) > 0 for |x| < 1 and B(x) = 0 for |x|  1,
• B(x) = B(−x),
• B(x) is non-increasing for x  0.
Let I = {1, 2, . . . , M} be the complete index set. Denote the kernel approximation of
p(z) with samples z j , for j ∈ I \ {i}, termed leave-one-out density, as
1 
pi (y|x) = κ((x, y)|z j ).
M − 1 j ∈I,j =i
1318 D Ruan

Conditioned on X = xi , one obtains a distribution of pi (y|xi ), and subsequently an

estimate ŷ i using any chosen estimate in section 2.3. Let ei = y i − ŷ i be the residual of the
observed sample response from the estimated response with the leave-one-out density. Let
ρ be a scale parameter, which could be set as the medium of |ei |’s for i = 1, 2, . . . , M. We
define the robust weight by
δi = B(y i − ŷ i /ρ). (15)
The original kernel density approximation (1) can then be modified by

1 
M
p(z) = M
δi κ(z|z i ) (16)
i=1 δi i=1

to incorporate the robust weighting.

2.5.4. Modified kernel weight to account for temporal correlation. By the same token as
in section 2.5.3, one could incorporate the temporal correlation between the training samples
and the test sample by modifying (16) further with

M
p(z) = δi ηi κ(z|z i ), (17)
i=1

where ηi is a monotonically decreasing function of the temporal distance between sample z i

and z. Fading memory is often modeled with exponential discounting or windowed training.
• For exponential discounting,
ηi = exp(−α|ti − t|), (18)
where ti and t are the time tags associated with z i and z, respectively. The positive
constant α determines the decay rate of influence of one sample on another as their
temporal distance increases.
• For moving window:

1 |t − ti | < ,
ηi = (19)
0 else,
where is the window size. Here, only training samples close enough in time to the test
sample are used to estimate the pdf. The window size should be chosen large enough
to ensure a reasonable kernel approximation of the pdf.

2.6. Benchmark methods for comparison

Sampling rate and lookahead length are the two major prediction parameters in adaptive image-
guided radiotherapy. We desire prediction algorithms that can handle low sampling rates and
large lookahead lengths: a low sampling rate means less imaging dose and a large lookahead
length allows more time for observation acquisition, signal processing and delivery. We will
study the performance of the proposed method, with varying sampling rates and lookahead
lengths and compare the outcome with the following benchmarks.
• Most recent sample (no prediction)
ŷ k = sk . (20)
Kernel density estimation-based real-time prediction for respiratory motion 1319

• Adaptive linear predictor

ŷ k = βkT xk + γk , (21)

where the linear coefficients βk and γk are obtained by solving the least-squares problem
for the observed covariate–response pairs in a dynamically updated training set. More
specifically, at each instant k, a training set of covariate–response pairs is constructed with
the most recent observed samples, and the prediction coefficients are determined by

(βˆk , γˆk ) = arg min (yi − βk xi − γk )2 ,
i∈ training set for time k

which can be solved in closed form.

3. Materials and results

3.1. Material

We used the real position management system (RPM system, Varian Medical, Palo Alto,
CA) to obtain 1D traces of fiducial markers placed on the patient’s chest wall. The RPM
traces are believed to be highly correlated with respiratory motion and sufficiently capture
the temporal behavior of respiration. Moreover, the performance of respiratory prediction
algorithms depends on the fundamental variation pattern rather than the amplitude, so the
RPM traces are reasonable test subjects for algorithmic development.
To rid the adverse impact of the arbitrary scaling in RPM amplitude, we adopt the
normalized root mean squared error (nRMSE) as the performance measure for each trace,
defined by the usual RMSE divided by the standard deviation of the observed sample values:

 N  RMSE E((y − ŷ)2 )
nRMSE {ŷ}i=1 = = . (22)
stdy E((y − ȳ)2 )

Population nRMSE (across traces) is computed by taking the L2 average of the trace-wise
nRMSE, i.e.,

1 
nRMSE = nRMSE2i .
number of traces i:trace id

We report the RPM data characteristics in table 1 and illustrate some traces in figure 2.

Table 1. RPM dataset information.

Subject ID 1 2 3 4 5 6 7 8 9 10

STD 0.49 0.50 0.30 0.20 0.32 0.59 0.07 0.23 0.28 0.11
P-P 2.54 2.36 1.27 1.12 1.87 0.97 0.29 0.88 1.22 0.43
P-P/STD 5.11 4.74 4.21 5.66 5.92 5.61 4.59 3.88 4.44 4.05
Duration (s) 140 79 113 165 165 117 150 165 160 162
1320 D Ruan

0.5 0.5
1

0.5
0

0
0

−0.5 −0.5

−1 −0.5
−1
−1.5

−2 −1 −1.5
0 50 100 150 20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160

trace 1 trace 4 trace 5

0.1 0.5 0.2

0.05

0
0 0
−0.05

−0.1
−0.5 −0.2
−0.15

−0.2

−0.25 −1 −0.4
0 50 100 150 20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160

trace 7 trace 9 trace 10

Figure 2. Typical RPM traces used in this study.

3.2. Experiment detail and results

We have chosen to use a three-dimensional augmented covariate variable, so that xi =
[si−2 , si− , si ]. This low-dimensional decision was based on the consideration that we
would like to obtain a reasonable kernel density estimation from limited samples and avoid the
‘curse of dimensionality’ in density estimation. This setup is also designed to ensure fairness
in performance comparison with the benchmark methods.
To properly capture the motion dynamics, we choose  to correspond to 0.4 s delay
between consecutive coordinates of the covariate x. In the baseline setup, we used a sampling
rate of 30 Hz, with  = 12. As mentioned in section 2.5.1, we are most interested in a
predictor that could perform well after a short training stage under low sampling conditions,
so we set the covariance in (3) with the estimated population covariance from the training
samples as in (12): this covariance setup corresponds to an overly broad (flat) local kernel, but
it provides numerical stability in most cases, an issue discussed in section 4 . We investigated
a lookahead length L = 30 corresponding to a 1 s prediction, which has been reported to be
challenging for a wide spectrum of common prediction techniques (Vedam et al 2003, Sharp
et al 2004, Murphy and Dieterich 2006).
For a fair comparison, the adaptive linear filter uses the same covariate variables, and
the observed covariate–response pairs in the most recent 20 s are used to estimate the linear
regression coefficients βk ∈ 3 and γk ∈  at each instant k.

3.2.1. Training samples in kernel density estimation. We studied three different schemes in
choosing the training samples used in kernel density estimation. In the static scheme, the first
20 s of the trajectory was used to generate the training sample collection, and the estimated
pdf was kept still thereafter for all predictions. In the expansive scheme, the training set is
enriched as new covariate–response pairs were observed. This corresponds to a special case
of fading memory (cf section 2.5.4) where all previous samples contribute equivocally to the
kernel approximation (M = k − L, α = 0 in (17) and (18)). In the moving window update
scheme, only training samples within the most recent 20 s temporal window were used to
Kernel density estimation-based real-time prediction for respiratory motion 1321

1 0.5

0.5
0
RPM displacement

RPM displacement
0

−0.5 −0.5

−1
−1
−1.5

−2 −1.5
0 20 40 60 80 100 120 140 20 40 60 80 100 120 140 160
time (second) time (second)

(a) trace 1 - major drift (b) trace 5 - transient changes

0.2
0.2

0
RPM displacement

RPM displacement
0
−0.2

−0.4

−0.6 −0.2

−0.8

−1 −0.4
20 40 60 80 100 120 140 160 20 40 60 80 100 120 140
time (second) time (second)

(c) trace 9 - irregular (d) trace 10 - quasi regular

Figure 3. Comparison among 1 s lookahead prediction results with different training schemes
in kernel density estimation. Actual signal trajectory (solid blue line), prediction with the static
training scheme (dashed green line), prediction from the expansive training scheme (dashed cyan
line), prediction from the moving window training scheme (dashed red line).

Table 2. Comparison of prediction performance among static training, expansive training and
moving window training.

Subject ID 1 2 3 4 5 6 7 8 9 10 Average

Root mean squared error (RMSE)

Static 1.85 0.81 0.73 0.91 0.99 1.03 0.83 0.54 1.05 0.85 1.02
Expansive 0.58 0.49 0.48 0.64 0.56 0.54 0.72 0.41 0.72 0.63 0.59
Moving window 0.37 0.37 0.39 0.58 0.45 0.42 0.69 0.35 0.60 0.56 0.49

construct the kernel pdf, as formulated in (19). Table 2 reports the prediction performance of
these three training schemes and figure 3 illustrates some typical prediction trajectories with
these training schemes.
It is quite obvious that expanding the training samples, or even better, using a moving
window to select the training sample collection, improves the prediction performance. This
is no surprise, as the updated pdf would drive the final prediction estimate to resemble the
training samples that both behave similarly and are close in time. This is particularly true for
trajectories with drifting, as shown in figure 4.
1322 D Ruan

0.5

RPM displacement
0

−0.5

−1

−1.5

−2
0 20 40 60 80 100 120 140
time (second)

(a) trajectory with trendy mean drift (patient 1)

0.04 0.05 0.08

0.035 0.07
0.04
0.03 0.06

0.025 0.03 0.05

0.02 0.04

0.015 0.02 0.03

0.01 0.02
0.01
0.005 0.01

0 0 0
−1 −0.5 0 0.5 1 1.5 2 2.5 −1 −0.5 0 0.5 1 1.5 −1 −0.5 0 0.5 1

(b) static training (c) expansive training (d) moving-window training

Figure 4. Comparison among 1 s lookahead prediction results with different training samples in
kernel density estimation for drifting trajectory. Top row: (a) time series of actual signal trajectory
(solid blue line), prediction with the static training scheme (dashed green line), prediction from
the expansive training scheme (dashed cyan line), prediction from the moving window training
scheme (dashed red line). Bottom row: histogram of prediction residual yk − ŷk for (b) static
training; (c) expansive training ; (d) moving window training.

3.2.2. The effect of sampling rate and lookahead length. Since sampling rate and lookahead
length are two of the most critical parameters in a prediction system, we compared the
performance of the proposed method with the most recent sample and the adaptive linear
filter prediction method as described in section 2.6, for various sampling rates and lookahead
lengths. In particular, we tested prediction performances for lookahead lengths 0.2 s, 0.6 s
and 1 s when samples are acquired at 5 Hz, 10 Hz, 15 Hz and 30 Hz respectively. Figure 5
reports the nRMSE for the three methods (MRS, Linear, Kernel) under different combinations
of prediction parameters. In figure 5, the upper-left corner corresponds to the relatively easy
case of short prediction (0.2 s) with dense samples (30 Hz). The sampling rate decreases as
we move down the rows and the prediction length increases as we move to the columns on the
right—both changes increases the difficulty of prediction.
In most cases, the adaptive linear filter yields significant accuracy gain compared to the
MRS, usually reducing the nRMSE by about a half. This performance is comparable to what
is reported by current development of respiratory predictors, justifying our previous argument
that the adaptive linear predictor is a fair benchmark for this study. The only exception occurs
when 5 Hz samples are used to predict 0.2 s ahead. This can be explained from two aspects:
(1) 200 ms delay is relatively short and does not incur too high an nRMSE even using MRS;
(2) the low sampling rate basically renders a down-sampled path, causing relative slow response
of the adaptive linear filter to dynamic changes. Meanwhile, the proposed kernel-based method
universally dominates the alternatives on a trace-to-trace basis, always yielding the lowest
nRMSE value. Moreover, its improvement upon RMS and adaptive linear predictor is quite
significant.
Kernel density estimation-based real-time prediction for respiratory motion 1323

0.4 1.2 2

0.35
MRS 1 MRS MRS
0.3 1.5
Linear Linear Linear
0.25 Kernel Kernel Kernel
0.8

nRMSE

nRMSE
nRMSE

0.2 1
0.6
0.15

0.1 0.5
0.4
0.05

0 0.2 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Trace ID Trace ID Trace ID

sampling at 30Hz, lookahead (left to right) 0.2s, 0.6s, 1s.

0.5 1.2 2

0.45
MRS 1 MRS MRS
0.4 1.5
Linear Linear Linear
0.35 Kernel Kernel Kernel
0.8
nRMSE

nRMSE
nRMSE

0.3 1
0.6
0.25

0.2 0.5
0.4
0.15

0.1 0.2 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Trace ID Trace ID Trace ID

sampling at 15Hz, lookahead (left to right) 0.2s, 0.6s, 1s.

0.5 1.2 2

0.45
MRS 1 MRS MRS
0.4 1.5
Linear Linear Linear
0.35 Kernel Kernel Kernel
0.8
nRMSE

nRMSE
nRMSE

0.3 1
0.6
0.25

0.2 0.5
0.4
0.15

0.1 0.2 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Trace ID Trace ID Trace ID

sampling at 10Hz, lookahead (left to right) 0.2s, 0.6s, 1s.

0.6 1.2 2

0.5 MRS 1 MRS MRS

1.5
Linear Linear Linear
Kernel Kernel Kernel
0.4 0.8
nRMSE

nRMSE

nRMSE

1
0.3 0.6

0.5
0.2 0.4

0.1 0.2 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Trace ID Trace ID Trace ID

sampling at 5Hz, lookahead (left to right) 0.2s, 0.6s, 1s.

Figure 5. Comparison of prediction performance in terms of nRMSE under various prediction

parameters. Column-wise (left → right): lookahead = 0.2 s, 0.6 s, 1 s. Row-wise (up → down):
sampling rate = 30 Hz, 15 Hz, 10 Hz, 5Hz.

Figure 6 reports the nRMSE across all traces with varying sampling rates for lookahead
lengths 0.2 s, 0.6 s and 1 s respectively, and figure 7 reports the performance subject to
varying lookahead lengths, for each given sampling rate. The advantage of the proposed
method is universal across all prediction lengths and all sampling rates and is most dramatic
for long lookahead length when the alternative methods yield essentially useless prediction
(nRMSE  1). The nonparametric nature and data-driven learning makes the kernel-based
prediction robust toward the changes in the prediction parameters.
1324 D Ruan

0.5 1 1.3

1.2
0.45 0.9
MRS MRS 1.1 MRS
0.4 Linear 0.8 Linear Linear
1
Kernel Kernel Kernel
0.7

nRMSE

nRMSE
0.35 0.9
nRMSE

0.6 0.8
0.3
0.7
0.25 0.5
0.6
0.2 0.4
0.5

0.4
5 10 15 20 25 30 5 10 15 20 25 30 5 10 15 20 25 30
sampling frequency (Hz) sampling frequency (Hz) sampling frequency (Hz)

(a) lookahead 0.2 s (b) lookahead 0.6 s (c) lookahead 1 s

Figure 6. nRMSE versus sampling frequency for 0.2, 0.6 and 1 s lookahead prediction across all
traces.

1.6 1.6

1.4 1.4
MRS MRS
1.2 Linear 1.2 Linear
Kernel Kernel
1 1
nRMSE

nRMSE

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
lookahead length (seconds) lookahead length (seconds)

(a) sampling frequency 5 Hz (b) sampling frequency 10Hz

1.4 1.4

1.2 1.2
MRS MRS
1 Linear 1 Linear
Kernel Kernel
0.8 0.8
nRMSE

nRMSE

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
lookahead length (seconds) lookahead length (seconds)

(c) sampling frequency 15 Hz (d) sampling frequency 30 Hz

Figure 7. nRMSE versus lookahead length with various sampling rates across all traces.

4. Discussions

• The Gaussian mixture structure in (6) offers an efficient form to store and process the pdf.
It suffices to record the weights wi , the Gaussian centers yi and the variance σy to fully
recover the probability distribution. Furthermore, the Gaussian kernel with separable
Kernel density estimation-based real-time prediction for respiratory motion 1325

covariance (with respect to covariate and response variables), together with the linearity
in the mean estimate allows us to obtain the prediction ŷ mean as a simple weighted sum
of sample response values (8), where the weight of each training sample is determined by
how close its covariate value is to the covariate value of the test sample.
• When there are only a few training samples (with respect to the covariate–response space
under study) available, numerical instability may occur if each of the weights wi in (5)
is small relative to the machine precision. A ‘zero divided by zero’ fault may arise from
normalization. Geometrically, uniformly small weights indicate that the test sample is far
from all training samples. One could utilize this observation to detect ‘unseen’ changes
in the trajectory. From the perspective of obtaining a stable estimate, one could adjust
the covariance x in (5) to ‘flatten’ the local kernel and increasing its bandwidth, so the
support of the kernel density approximation from the same training samples covers a larger
portion of the whole space, preventing uniformly vanishing p(y|x). As an alternative,
one could use a finitely supported density kernel and determine a test-sample-dependent
bandwidth by requesting the test sample to fall inside the kernel support of a certain
number of training samples, as in Cleveland (1979), Silverman (1986), Ruan et al (2007).

5. Conclusion and future work

This paper reports a kernel density-based method to estimate the probability distribution of
the future target position. We provide a general framework for estimating the conditional
probability distribution and several options in constructing estimates based on the conditional
distribution. In particular, we have discussed the use of the Gaussian density kernel and
the mean estimate, leading to a simple yet intuitive estimate that is the weighted sum of
the training response values. Our proposed method compares favorably with alternative
benchmark methods, yielding a reduction of two- to threefolds in RMSE. Improvement of the
proposed method is most noticeable in the case of low sampling rate and/or long lookahead
length prediction.
We have discussed how numerical instability could be indicative of an ‘unseen’ scenario,
and may be used for change detection. More generally, the marginal distribution of the test
covariate p(x) indicates the chances of such event and the conditional distribution p(y|x)
provides all the information for characterizing the quality of the resulting estimate. This gives
rise to an instantaneous quantification of the prediction, which is an interesting alternative to the
conventional collective performance measure. We will study the performance quantification
issue as a natural extension of this work.
As we described in our methodology and observed in our experiment, the choice of the
covariance for the kernel density is not particularly critical for the purpose of prediction.
However, a good estimation of the joint distribution of the covariate–response is valuable for
its own sake. Related to the previous comment, change detection relies on a good estimate of
the marginal distribution. We will study automatic schemes to ‘optimally’ choose the kernel
covariance.
We have formulated the mode and medium estimates in section 2.3, but have focused
on the mean estimate due to its simplicity and the resulting intuitive weighted superposition
form for the predictor. However, the mode estimate corresponds to the maximum a posteriori
(MAP) estimate and the medium estimate has its advantage toward outlier samples, and have
their own significance. In the future, we will analyze these estimates and investigate their
applicability in various settings.
1326 D Ruan

Acknowledgments

The author thanks Dr Per Poulsen and Dr Byung-Chul Cho for motivating this work and
Dr Paul Keall for his support. She is grateful to the reviewers for their insightful comments
that greatly improved the quality of this work.

References

Botev Z I 2006 A novel nonparametric density estimator Postgraduate Seminar Series Department of Mathematics,
The University of Queensland
Cleveland W S 1979 Robust locally weighted regression and smoothing scatterplots J. Am. Stat. Assoc. 74 829–36
Duda R O, Hart P E and Stork D G 2001 Pattern Classification (New York: Wiley)
Ernst F, Schlaefer A and Schweikard A 2007 Prediction of respiratory motion with wavelet-based multiscale
autoregression Med. Image Comput. Comput. Assist. Intervention 10 668–75
Isaksson M, Jalden J and Murphy M J 2005 On using an adaptive neural network to predict lung tumor motion during
respiration for radiotherapy applications Med. Phys. 32 3801–9
Kakar M, Nystrom H, Aarup L R, Nottrup T J and Olsen D R 2005 Respiratory motion prediction by using the
adaptive neuro fuzzy inference system (ANFIS) Phys. Med. Biol. 50 4721–8
Keall P J, Kini V R, Vedam S S and Mohan R 2002 Potential radiotherapy improvements with respiratory gating
Australas Phys. Eng. Sci. Med. 25 1–6
McCall K C and Jeraj R 2007 Dual-component model of respiratory motion based on the periodic autoregressive
moving average (periodic ARMA) method Phys. Med. Biol. 52 3455–66
McMahon R, Papiez L and Sandison G 2007 Addressing relative motion of tumors and normal tissue during dynamic
MLC tracking delivery Australas Phys. Eng. Sci. Med. 30 331–6
Murphy M J and Dieterich S 2006 Comparative performance of linear and nonlinear neural networks to predict
irregular breathing Phys. Med. Biol. 51 5903–14
Nuyttens J J, Prevost J B, Praag J, Hoogeman M, Van Klaveren R J, Levendag P C and Pattynama P M 2006 Lung
tumor tracking during stereotactic radiotherapy treatment with the cyberknife: marker placement and early
results Acta Oncol. 45 961–5
Putra D, Haas O C, Mills J A and Burnham K J 2008 A multiple model approach to respiratory motion prediction for
real-time IGRT Phys. Med. Biol. 53 1651–63
Ruan D, Fessler J A and Balter J M 2007 Real-time prediction of respiratory motion based on nonparametric local
regression methods Phys. Med. Biol. 52 7137–52
Sharp G C, Jiang S B, Shimizu S and Shirato H 2004 Prediction of respiratory tumour motion for real-time image-
guided radiotherapy Phys. Med. Biol. 49 425–40
Sheather S J and Jones M C 1991 A reliable data-based bandwidth selection method for kernel density estimation J.
R. Stat. Soc. B 53 683–90
Silverman B W 1986 Density Estimation for Statistics and Data Analysis (New York: Chapman and Hall)
Vedam S S, Kini V R, Keall P J, Ramakrishnan V, Mostafavi H and Mohan R 2003 Quantifying the predictability of
diaphragm motion during respiration with a noninvasive external marker Med. Phys. 30 505–13
Vedam S S, Keall P J, Docef A, Todor D A, Kini V R and Mohan R 2004 Predicting respiratory motion for
four-dimensional radiotherapy Med. Phys. 31 2274–83
Zimmerman J, Korreman S, Persson G, Cattell H, Svatos M, Sawant A, Venkat R, Carlson D and Keall P 2008
DMLC motion tracking of moving targets for intensity modulated arc therapy treatment—a feasibility study
Acta Oncol. 1–6