IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 4, JULY 2015 1271 Toward Noninvasive Quantification of Brain Radioligand Binding by Combining Electronic Health Records and Dynamic PET Imaging Data Arthur Mikhno, Francesca Zanderigo, R. Todd Ogden, J. John Mann, Elsa D. Angelini, Andrew F. Laine, and Ramin V. Parsey Abstract—Quantitative analysis of positron emission tomogra- analysis of PET data requires both the arterial input function phy (PET) brain imaging data requires a metabolite-corrected ar- (AIF) that describes the amount of radioligand available for dif- terial input function (AIF) for estimation of distribution volume fusion into the brain, and the tissue time-activity curves (TACs) and related outcome measures. Collecting arterial blood samples adds risk, cost, measurement error, and patient discomfort to PET derived from dynamic PET images. Kinetic modeling is then studies. Minimally invasive AIF estimation is possible with simul- performed to estimate important outcome measures of radioli- taneous estimation (SIME), but at least one arterial blood sample is gand distribution and binding [1]. To measure the AIF, typically necessary. In this study, we describe a noninvasive SIME (nSIME) a catheter is inserted into the radial artery at a subjects’ wrist approach that utilizes a pharmacokinetic input function model to sample blood for the duration of the PET scan. After cen- and constraints derived from machine learning applied to an elec- tronic health record database consisting of “long tail” data (digital trifugation, the total radioactivity concentration of radioligand records, paper charts, and handwritten notes) that were collected in the arterial plasma (TP) is measured in each blood sample. If ancillary to the PET studies. We evaluated the performance of the body metabolizes the radioligand, the parent fraction (PF) nSIME on 95 [11C]DASB PET scans that had measured AIFs. The of unmetabolized radioligand in the plasma is assayed from a results indicate that nSIME is a promising alternative to invasive subset of the blood samples. After fitting PF using a metabolite AIF measurement. The general framework presented here may be × expanded to other metabolized radioligands, potentially enabling model, the input function is calculated as AIF : y =TP PF, quantitative analysis of PET studies without blood sampling. A which reflects the concentration y of radioligand in plasma that glossary of technical abbreviations is provided at the end of this is available to enter the target tissue. paper. Utilizing the AIF and the TACs, PET imaging can be used Index Terms—Arterial input function (AIF), electronic health to estimate outcome measures related to the “binding poten- record (EHR), positron emission tomography (PET) imaging. tial” of a radioligand to its target. In particular, one estimate of binding potential BPF is defined as: BPF = Bavail/KD = I. INTRODUCTION (VT − VND)/fP , where Bavail is the concentration of available OSITRON emission tomography (PET) uses radioactively receptors, 1/KD is the radioligand affinity to the target, VT P tagged probes (radioligands) for the in vivo quantification is the radioligand “volume of distribution” or volume of ra- of blood flow, metabolism, protein distribution, gene expres- dioligand in tissue relative to plasma, VND is the radioligand sion, and drug target occupancy in the brain. Fully quantitative “volume of distribution” in a tissue devoid of the target (i.e., fraction of binding not specific to the target of interest), and fP Manuscript received October 15, 2014; revised January 5, 2015 and March is the free fraction of the radioligand in plasma. VT and VND 6, 2015; accepted March 14, 2015. Date of publication March 24, 2015; date of are estimated from kinetic modeling of the region TACs and the current version July 23, 2015. This work was supported by the National Insti- tute of Mental Health (NIMH) F31 Predoctoral Fellowship (1F31MH095338) AIF, while fP can be assayed in additional blood samples col- and NIMH Grants MH040695 and MH062185, and the National Center for lected prior to radioligand injection. When fP is not available, Advancing Translational Sciences, National Institutes of Health, Grant TL1 or cannot be measured reliably, two other variants binding po- TR000082. − A. Mikhno, E. D. Angelini, and A. F. Laine are with the Department of tential can be calculated: BPP = fP BPF =(VT VND) and Biomedical Engineering, Columbia University, New York, NY 10027 USA BPND = f NDBPF =(V T − V ND)/V ND, where fND is the (e-mail: [email protected]; [email protected]; AL418@columbia. free fraction of the radioligand in a tissue devoid of target. edu). F. Zanderigo is with the Department of Psychiatry, Columbia Univer- Thus, quantification of PET data requires arterial blood sam- sity and the Division of Molecular Imaging and Neuropathology, New pling to estimate the AIF and to calculate the outcome measures York Psychiatric Institute, New York, NY 10021 USA (e-mail: zanderi@ related to the “binding potential” of the radioligand to its target nyspi.columbia.edu). R. Todd Ogden is with the Department of Biostatistics, Columbia University, (i.e., BPP or BPND). While arterial sampling is routinely done New York, NY 10032 USA (e-mail: [email protected]). in research studies, it is invasive, necessitates specific techni- J. J. Mann is with the Department of Psychiatry, Columbia University, New cal expertise, exposes clinical personnel to radiation, involves York, NY 10032 USA (e-mail: [email protected]). R. V. Parsey is with the Department of Psychiatry, Stony Brook University, laboratory analysis costs, significant measurement error, and Stony Brook, NY 11794 USA (e-mail: Ramin.Parsey@stonybrookmedicine. strongly discourages subject participation in PET studies. If the edu). patient refuses an arterial line, or if arterial cannulation or blood Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. assay fails, the entire study may be dropped from data analy- Digital Object Identifier 10.1109/JBHI.2015.2416251 sis leaving expensive PET images that cannot be interpreted or 2168-2194 © 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information. 1272 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 19, NO. 4, JULY 2015 analyzed. Even when these challenges are anticipated, it may not ROIs. SIME exploits a parametric model for the unknown AIF always be possible to perform arterial blood sampling, such as and seeks to estimate model parameters simultaneously with the with vulnerable populations (e.g., elderly, cancer patients), those kinetic parameters related to the binding, while fitting several with movement disorders (e.g., people affected by Parkinson’s ROIs at the same time. Recent work has shown that SIME can disease) or in a combined magnetic resonance imaging (MRI) recover the AIF using only a single arterial blood sample as and PET scanner due to magnetic interference. Even when blood an anchor (or constraint) that ensures model identifiability [15]. samples can be obtained, fitting of TP and PF can be challeng- One more alternative for a “less” invasive IDIF is to measure TP ing due to inherent noise present in the blood measurements, and PF from venous blood sampling [3]. Drawing venous blood and complex radioligand kinetics such as “lung trapping” that does not require arterial catheterization by trained personnel, require adapting and validating a metabolite model [2]. Finally, making it more practical in research and clinical settings. For and most commonly, arterial sampling is impractical in a clin- some radioligands, this substitution is nearly equivalent as with ical setting, hindering adoption of quantitative PET outcome [18F]FDG [10], [16] and [11C]WAY [17]. The equivalence be- measures for clinical use. tween arterial and venous blood must be determined separately The last decade has brought a considerable effort to develop for each radioligand and often does not hold. This procedure AIF estimation techniques [3] that can be broadly categorized still requires considerable effort with regards to blood draws as noninvasive (i.e., without blood samples), such as “reference and the metabolite correction assay. tissue” and population-based approaches, and minimally inva- A totally noninvasive AIF estimation approach is needed to sive (i.e., with few blood samples), such as image-derived input overcome the aforesaid challenges and drawbacks with existing functions (IDIF). IDIF techniques. Our proposed solution taps into the field of Reference tissue methods use only imaging data to estimate population pharmacokinetics and pharmacodynamics (PPKD) kinetic parameters of tissue TACs, based on the specification that focuses on predicting metabolite corrected blood levels of that a valid reference region devoid of the target (e.g., receptor pharmaceutical compounds at various time points after injec- or protein of interest) exists and can be identified [4]. This ap- tion. This is done by aggregating blood data from many subjects proach allows estimation of BPND only. For many radioligands to determine what measures (e.g., age, body mass index (BMI), currently employed in brain studies, a reference region truly de- glomerular filtration rate, etc.) explain the variance in drug blood void of the target cannot be identified or the region commonly concentration [18]. Previous work suggests that combining sup- used as a reference actually has measurable specific binding. plementary information (e.g., weight, height, ID) with signal Even when a reference region is identifiable, there may be high from cranial blood vessels improves IDIF-based AIF estimation bias and variance in BPND values calculated with the reference [19]. However, this technique was developed for the radioligand tissue methods when compared to using the AIF, as is the case [18F]FDG for which metabolites are not present in the blood with [11C]-PK11195 (targeting microglia) [5], and [11C]ABP (i.e., PF = 1).