Thesis (Ph. D.)--Harvard-MIT Division of Health Sciences and Technology, 2004. / Includes bibliographical references (p. 49-51). / Text documents generally contain two forms of structures, logical structures and physical structures. Loosely speaking, logical structures are sections of text that are both visually and semantically distinct. For example, a document may have an "introduction", a "body", and a "conclusion" as its logical structures. These structures are so named because each section has a distinct purpose in conveying the document's logical arguments or intentions. Perfect machine recognition of logical structures in large collections of documents is an unsolved problem in computational linguistics. This thesis presents evidence that a new family of functions on text segments carries information that is useful for differentiating document logical structures. For any given text segment, a function of this form is referred to as the cadence, and it is based on a new interpretation of the vector space representation that Gerard Salton introduced in 1975. Cadence also differs from the original Salton representation in that it relies on three heuristic transformations based on authorship, location, and term coherence. To test the hypothesis that the cadence of a text segment carries information helpful to differentiating logical structures, a corpus was built containing 2800 documents with manually-annotated logical structures. Structures representing abstracts, introductions, bodies, and conclusions from this corpus were clustered with a k-means algorithm using cadence data. Precision and recall performances were computed for the results, and a chi-squared cross-tabulation test was used to determine the statistical significance of the clustering results. Precision and recall were highest for abstracts (P = 0.931 [plus-minus] 0.025, R = 0.992 / (cont.) [plus-minus] 0.026), followed by introductions (P = 0.747 [plus-minus] 0.025, R = 0.802 [plus-minus] 0.026) and conclusions (P = 0.737 [plus-minus] 0.025, R = 0.813 [plus-minus] 0.026), and lowest for bodies (P = 0.876 [plus-minus] 0.03, R = 0.663 [plus-minus] 0.026). These results suggest that cadence may have substantial promise for finding logical structures in un-annotated documents. / by Ngon D. Dao. / Ph.D.

Thesis (S.M.)--Harvard-MIT Division of Health Sciences and Technology, 2011. / Cataloged from PDF version of thesis. / Includes bibliographical references (p. 50-53). / Founders of new ventures in the medical technology (Medtech) industry require capital to establish, sustain, and grow their companies. Most founders must seek some form of external capital to meet these demands; in Medtech, the most well-known and prestigious of these is venture capital (VC). However, another type, angel investors, may be as important as VCs. Angels are accredited investors that invest their own money directly in new ventures. Founders of new Medtech ventures may choose to seek capital from angel investors in addition to, or instead of, venture capitalists. Unfortunately, there is little research available on outcomes for founders and their firms when angel investors are involved. Like VCs, angels seek financial returns from their investments; however, there may be additional and different motivations at play that make angels willing to grant more friendly terms to founders. As a result, it may actually be advantageous for founders to seek capital from angel investors. This paper addresses the question of whether founders of new ventures in the Medtech industry have better outcomes in terms of ownership and control of the company when one or more investment rounds involve angel investors in addition to, or in place of, VCs. Ownership is measured by the amount of equity owned just prior to an IPO, and control by the presence of founders as employees or directors at the time of the IPO. Analyzing S-is from the last 10 years of initial public offerings (IPOs), a dataset was constructed that comprised the shareholders of the 63 Medtech companies that experienced an IPO between 2001 and 2010. Of these, 18 companies had some presence of angel ownership that could be gleaned from the S-1; of those, 12 had at least a 5% stake belonging to angels. Results presented in the paper show, for the first time, those founders of Medtech firms with angel investors as shareholders at the time of IPO have significantly greater ownership of shares and significantly greater control of the firm as an employee or director than founders of firms without angels present. Angel-backed firms required less investment capital and no more time to reach the IPO, and, importantly, did not suffer with respect to the overall valuation of the firm. On the contrary, there was a trend of firms - and founders themselves - seeming to benefit from a valuation perspective, and significantly better from a multiple perspective, when angel investors were present. Even when firms received backing from venture capitalists, angel investor involvement also seemed to generally improve the performance of the firm and of the founders along the measured dimensions. / by Alan R. Braly. / S.M.

Thesis (Ph. D.)--Harvard-MIT Division of Health Sciences and Technology, 2007. / This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. / Includes bibliographical references (p. 211-225). / While cochlear implantation has become the standard care in treating patients with severe to profound sensorineural hearing loss, the variation in benefit (communicative ability) individual patients derive from implantation remains both large and, for the most part, unexplained. One explanation for this variation is the status of the implanted ear which, when examined histopathologically, also displays substantial variation due to both the pathogenesis of hearing loss (etiology, etc.) and pathological changes initiated by implantation. For instance, across-patient variation in electrode position and insertion depth is clearly present, as are differential amounts of residual spiral ganglion survival, fibrous tissue formation and electrode encapsulation, cochlear ossification, and idiosyncratic damage to adjacent cochlear structures. Because of the complex geometric electrical properties of the tissues found in the implanted ear, demonstrating the impact of pathological variability on neuronal excitation, and ultimately on behavioral performance, will likely require a detailed representation of the peripheral anatomy. Our approach has been to develop detailed, three-dimensional (3D) electro-anatomical models (EAMs) of the implanted ear capable of representing the aforementioned patient-specific types of pathological variation. In response to electric stimulation, these computational models predict an estimate of (1) the 3D electric field, (2) the cochleotopic pattern of neural activation, and (3) the electrically-evoked compound action potential (ECAP) recorded from intracochlear electrodes. This thesis focuses on three aims. First, two patient-specific EAMs are formulated from hundreds of digital images of the histologically-sectioned temporal bones of two patients, attempting to incorporate the detailed pathology of each. Second, model predictions are compared to relevant reports from the literature, data collected from a cohort of implanted research subjects, and, most importantly, to archival data collected during life from the same two patients used to derive our psychophysical threshold measures, and ECAP recordings) collectively show a promising correspondence between model-predicted and empirically-measured data. Third, by making incremental adjustments to the anatomical representation in the model, the impact of individual attributes are investigated, mechanisms that may degrade benefit suggested, and potential interventions explored. / by Darren M. Whiten. / Ph.D.

Thesis (Ph. D.)--Harvard-MIT Division of Health Sciences and Technology, 2007. / Includes bibliographical references (leaves 126-142). / Complex traits are conditions that, as a result of the complex interplay among genetic and environmental factors, have wide variability in progression and manifestation. Because most common diseases with high morbidity and mortality are complex traits, uncovering the genetic architecture of these traits is an important health problem. Asthma, a chronic inflammatory airway disease, is one such trait that affects over 300 million people around the world. Although there is a large amount of human genetic information currently available and expanding at a rapid pace, traditional genetic studies have not provided a concomitant understanding of complex traits, including asthma and its related phenotypes. Despite the intricate genetic background underlying complex traits, most traditional genetic studies focus on individual genetic variants. New methods that consider multiple genetic variants are needed in order to accelerate the understanding of complex traits. In this thesis, the need for better analytic approaches for the study of complex traits is addressed with the creation of a novel method. Probabilistic graphical models (PGMs) are a powerful technique that can overcome limitations of conventional association study approaches. / (cont.) Going beyond single or pairwise gene interactions with a phenotype, PGMs are able to account for complex gene interactions and make predictions of a phenotype. Most PGMs have limited scalability with large genetic datasets. Here, a procedure called phenocentric Bayesian networks that is tailored for the discovery of complex multivariate models for a trait using large genomic datasets is presented. Resulting models can be used to predict outcomes of a phenotype, which allows for meaningful validation and potential applicability in a clinical setting. The utility of phenocentric Bayesian networks is demonstrated with the creation of predictive models for two complex traits related to asthma management: exacerbation and bronchodilator response. The good predictive accuracy of each model is established and shown to be superior to single gene analysis. The results of this work demonstrate the promise of using the phenocentric Bayesian networks to study the genetic architecture of complex traits, and the utility of multigenic predictive methods compared to traditional single-gene approaches. / by Blanca Elena Himes. / Ph.D.

Thesis: Ph. D., Harvard-MIT Program in Health Sciences and Technology, 2015. / Cataloged from PDF version of thesis. / Includes bibliographical references. / Vertebral fractures are the most common complication of osteoporosis and are associated with significant pain, height loss, disfigurement, respiratory impairment, depression, and decreased life span. Despite the high personal and societal costs of vertebral fractures, little is known regarding their biomechanical etiology. In particular, whereas much is known about the determinants of vertebral strength, little is known about the in vivo loading of the spine that may contribute to vertebral fracture. Prior efforts to understand the possible contribution of spine mechanics to vertebral fractures have been limited by the inability to accurately assess in vivo spinal loading, especially in the thoracic region. Thus, the overall goal of this work was to improve the understanding of vertebral fractures through detailed analysis of spinal loading. We first developed and validated a novel musculoskeletal model capable of predicting forces in the thoracolumbar spine during daily activities. Model-derived predictions of vertebral compressive loading and trunk muscle activity were highly correlated with previously collected in vivo measurements of pressure, vertebral compression from telemeterized implants, and trunk muscle myoelectric activity from electromyography. To gain insights into how individual variation in trunk anatomy influences vertebral loading, we developed a robust set of methods for rapid, automated generation of subject-specific musculoskeletal models of the thoracolumbar spine using computed tomography based measurements of spine curvature and trunk muscle morphology. Using these subject-specific models, we found that normal variations in spine curvature and muscle morphology in the adult population have a large effect on vertebral loading predictions. Specifically, we found that increasing thoracic kyphosis and reducing lumbar lordosis, changes that commonly occur with age, were both associated with higher spinal loads. Lastly, we used our musculoskeletal model to describe how vertebral loading and the factor-of-risk (load-to-strength ratio) vary along the spine for a large number of activities. For a majority of activities, the highest loads and factor-of- risk were in the thoracolumbar region, which is the spine region with the highest incidence of vertebral fracture. Further, we identified a unique biomechanical mechanism responsible for the high loads in this region. / by Alexander G. Bruno. / Ph. D.

Thesis: Ph. D., Harvard-MIT Program in Health Sciences and Technology, 2014. / Cataloged from PDF version of thesis. / Includes bibliographical references. / Biofilms-surface-bound communities of microbes-are a major medical concern, as they can be sources of infection that are difficult to eradicate. Their formation starts with the attachment of bacteria to available surfaces-often implantable biomaterials. The development of materials that prevent bacterial adhesion is therefore of paramount importance, and it requires a thorough understanding of the materials and bacterial surface properties that enable adhesive interactions. We herein design model surfaces and examine the interplay between micro-scale geometry, surface energy and bacterial surface properties with respect to adhesion, with the ultimate goal of understanding bacterial adhesion in structured environments, and establishing principles for design of novel surfaces that effectively repel bacteria. We first study adhesion of Escherichia coli to engineered surfaces possessing superficially unfavorable geometries. We show that cells can overcome geometric constraints with the aid of flagella, which are able to reach between narrow crevices, thus improving adhesion and expanding the range of surfaces to which cells can adhere. We examine binding of purified flagella to abiotic surfaces by means of quartz crystal microbalance (QCM) and show that flagella bind preferentially to hydrophobic surfaces, yet they do not appreciably bind to hydrophilic surfaces. Using mutant strains, we investigate the role of flagella in surface attachment of live cells and demonstrate that flagellated cells adhere best to hydrophobic substrates; however flagella may impede cell adhesion to hydrophilic surfaces. To further explore hydrophilic, structured environments with physiological relevance, we examine mucin-a natural hydrogel that typically harbors microbes in animals, while protecting the host. We purify mucins and use them in their native, three-dimensional configuration to probe bacterial swimming behavior and surface attachment in their presence. We demonstrate that mucins maintain-and possibly enhance-swimming ability for E. coli and Pseudomonas aeruginosa, and show that they greatly reduce adhesion to underlying substrates. Finally, we build on our established design principles and construct anti-adhesive surfaces by combining hydrophilic chemistries with topographic features smaller than cellular dimensions. This work suggests a path toward anti-adhesive materials that may be optimized for mechanical robustness, longevity and specific environments of application. / by Ronn S. Friedlander. / Ph. D.

Thesis (Ph. D.)--Harvard-MIT Division of Health Sciences and Technology, 2011. / Cataloged from PDF version of thesis. / Includes bibliographical references (p. 153-159). / Speech reception depends critically on temporal modulations in the amplitude envelope of the speech signal. Reverberation encountered in everyday environments can substantially attenuate these modulations. To assess the effect of reverberation on the neural coding of amplitude envelope, we recorded from single units in the inferior colliculus (IC) of unanesthetized rabbit using sinusoidally amplitude modulated broadband noise stimuli presented in simulated anechoic and reverberant environments. Consistent with the attenuation of amplitude modulation (AM) in the stimulus, both rate and temporal coding of AM were degraded in IC neurons. However, in most neurons, the degradation in temporal coding was smaller than the degradation in the stimulus. In many neurons, this compensation could be accounted for by the modulation input-output function (MIOF), which describes the nonlinear transformation of modulation depth from the sound stimulus into the neural response. However, in a subset of neurons, the MIOF underestimated the strength of temporal coding, suggesting that reverberant stimuli may have a coding advantage over anechoic stimuli with the same modulation depth. Additional experiments suggest that interaural envelope disparities and interaural decorrelation introduced by reverberation may partly explain this coding advantage. In another set of experiments, we tested the hypothesis that temporal coding of AM is not static, but depends dynamically on the modulation depth statistics of preceding stimulation. In a subset of neurons, preceding stimulation history significantly altered the MIOF. On average, temporal coding of modulation frequency was more robust in conditions when low modulation depths predominate, as in reverberant environments. Overall, our results suggest that the auditory system may possess mechanisms for reverberation compensation, and point to an important role of binaural and dynamic neural processes for robust coding of AM in reverberant environments. / by Michaël C. C. Slama. / Ph.D.

Thesis (Ph. D. in Biomedical Engineering)--Harvard-MIT Program in Health Sciences and Technology, 2012. / Cataloged from PDF version of thesis. / Includes bibliographical references (p. 142-151). / High throughput methods in molecular biology have changed the landscape of biomedical research. In particular, advances in massively parallel DNA sequencing and synthesis technologies are defining our genomes and the products they encode. In the first part of this thesis, we have constructed a rationally designed antibody library and analysis platform optimized for use with deep sequencing technologies. Libraries of fully defined oligonucleotides encode three complementarity determining regions (CDRs; L3 from the light chain, H2 and H3 from the heavy chain), and were combinatorially cloned into a synthetic single chain variable fragment (scFv) framework for molecular display. Our novel CDR sequence design utilized a hidden Markov model (HMM) that was trained on all antibody-antigen co-crystal complexes present in the Protein Data Bank. The resultant ~10¹² member library has been produced in ribosome display format, and was comprehensively analyzed over four rounds of antigen selections by multiplex paired-end Illumina sequencing. The HMM library generated multiple antibodies against an emerging cancer antigen and is the basis of a next generation antibody production platform. In a second application of these technologies, we have created a synthetic representation of the complete human proteome, which has been engineered for display on bacteriophage. We use this library together with deep DNA sequencing methods to profile the autoantibody repertoires of individuals with autoimmune disease in a procedure called phage immunoprecipitation sequencing (PhIP-Seq). In a proof-of-concept study, this method identified both known and novel autoantibodies contained in the spinal fluid of a control patient with paraneoplastic neurological syndrome. The study was then expanded to include a large scale automated screen of 289 independent antibody repertoires, including those from a large number of healthy donors, multiple sclerosis patients, rheumatoid arthritis patients, and type 1 diabetics. Our data describes each individual's unique "autoantibodyome", and defines a small set of recurrently targeted peptides in health and disease. / by Harry Benjamin Larman. / Ph.D.in Biomedical Engineering

Thesis (S.M.)--Harvard-MIT Division of Health Sciences and Technology, 2006. / Includes bibliographical references. / A well-designed implementation of medication concepts, records, and lists in an electronic medical record (EMR) system allows it to successfully perform many functions vital for the provision of quality health care. A controlled medication terminology provides the foundation for decision support services, such as duplication checking, allergy checking, and drug-drug interaction alerts. Clever modeling of medication records makes it easy to provide a history of any medication the patient is on and to generate the patient's medication list for any arbitrary point in time. Medication lists that distinguish between description and prescription and that are exportable in a standard format can play an essential role in medication reconciliation and contribute to the reduction of medication errors. At present, there is no general agreement on how to best implement medication concepts, records, and lists. The underlying implementation in an EMR often reflects the needs, culture, and history of both the developers and the local users. survey of a sample of medication terminologies (COSTAR Directory, the MDD, NDDF Plus, and RxNorm) and EMR implementations of medication records (OnCall, LMR, and the Benedum EMR) reveals the advantages and disadvantages of each. There is no medication system that would fit perfectly in every single context, but some features should strongly be considered in the development of any new system. / (cont.) A survey of a sample of medication terminologies (COSTAR Directory, the MDD, NDDF Plus, and RxNorm) and EMR implementations of medication records (OnCall, LMR, and the Benedum EMR) reveals the advantages and disadvantages of each. There is no medication system that would fit perfectly in every single context, but some features should strongly be considered in the development of any new system. / by Jaime Chang. / S.M.

Thesis (Ph. D. in Biomedical Engineering)--Harvard-MIT Program in Health Sciences and Technology, 2012. / Cataloged from PDF version of thesis. Vita / Includes bibliographical references. / We searched for genes that are potentially important for the maintenance of Pancreatic Ductal Adenocarcinoma (PDAC). PDAC is the 4th leading cause for cancer-related deaths and exhibits a 5-year survival rate of less than 5%. Since PDAC is a Kras-driven cancer in that greater than 90% of PDACs contain a Kras mutation, we tested genes that are downstream of Kras. We used RNAi technology to inhibit approximately 30 genes in the canonical Kras effector pathways, Mapk, Pi3k, and Ral. These genes were tested in the context of mouse cell lines derived from a genetically engineered mouse model of PDAC with conditional mutations in Kras G12D and p53. An individual gene-by-gene approach and a pooled high-throughput screening strategy were taken to identify important genes. We identified mTOR, and to a lesser extent, Raptor and Rictor, as genes that are important for the maintenance of PDAC both in vitro and in vivo. In addition, we show that inhibition of mTOR and Raptor are synthetic lethal in that PDAC lines are sensitive to their inhibition, while non-tumorigenic cell lines are not as sensitive. Moreover, inhibition of mTOR results in downregulation of mTORCl and mTORC2 targets, while inhibition of Raptor induces downregulation of only mTORC 1 targets. As combination therapies are likely to be more effective, we looked for a drug that could combine effectively with mTOR and Raptor. From screening several small molecule drugs that target the Mapk and Pi3k pathways, we found PDAC lines to be particularly sensitive to AZD6244, while normal cell lines are significantly less sensitive. The combination or either an mTOR or Raptor hairpin and AZD6244 was found to be additive in that the effect on viability is significantly greater than that of each intervention alone. Another approach to combinations is to combine two drugs. AZD6244 and BEZ235, an inhibitor of Pi3k and mTOR, were tested in combination on both PDAC and normal cell lines. The combination was synergistic in PDAC lines but not in normal lines, suggesting that the combination may be effective with low toxicity. In summary, through a screen, we have identified mTOR, Raptor, and Rictor, as being critical components for the maintenance of PDAC. / by Harry Jenq. / Ph.D.in Biomedical Engineering