Hematopathology / DATABASE FOR IMMUNOPHENOTYPING BY FLOW CYTOMETRY

A Relational Database for Diagnosis of Hematopoietic Neoplasms Using Immunophenotyping by Flow Cytometry
Andy N.D. Nguyen, MD,1 John D. Milam, MD,1 Kathy A. Johnson, PhD,2 and Eugenio I. Banez, MD3
Key Words: Hematopoietic neoplasms; Relational database; Immunophenotyping; Flow cytometry

Abstract
A relational database was developed to facilitate the diagnosis of hematopoietic neoplasms using results of immunophenotyping by flow cytometry. This database runs on personal computers and uses backwardchaining search to arrive at conclusions. Results of immunologic marker studies are processed by the database to obtain a set of differential diagnoses. The current version of this database includes diagnostic immunophenotyping pattern for 33 hematopoietic neoplasms. We tested this database using 92 clinical cases from 2 tertiary care medical centers. The database ranked the actual diagnosis as 1 of the top 5 differential diagnoses in 93% of the cases tested. The user can modify the database contents to suit individual needs. This database has been posted on the World Wide Web for direct access. We propose that this userfriendly database is a potential tool for computerassisted diagnosis of hematopoietic neoplasms.

Immunophenotyping has become one of the essential methods for proper classification of hematopoietic neoplasms. Flow cytometry used in immunophenotyping has no doubt added a new dimension to the diagnosis of leukemia and lymphoma.1,2 A wide range of monoclonal antibodies is available to recognize various hematopoietic cells based on their surface and cytoplasmic antigens.1-4 Leukemic and lymphoma cells usually cannot be detected with a single immunologic marker. Instead, the use of a monoclonal antibody panel consisting of multiple antibodies is required for supporting the provisional diagnosis based on histologic findings.1,2 Since many hematologic neoplasms demonstrate similar patterns of immunophenotyping,2-4 their diagnosis often presents a challenge to pathologists who interpret flow cytometric data. This is particularly applicable to practicing pathologists who are not subspecialized in hematopathology and to pathology residents-in-training. As the number of immunologic markers used in flow cytometry increases, a systematic approach for interpretation of marker results also is essential for consistent classification of neoplasms.5 The interpretation of immunophenotyping results by flow cytometry involves pattern recognition of different hematopoietic neoplasms that may have similar immunologic marker patterns. Each neoplasm is associated with a particular pattern characterized by the presence of certain markers and the absence of others.2-4 The numerous markers available in the flow cytometry laboratory make these patterns difficult to remember, especially for those of uncommon neoplasms.24 Another factor that hinders the interpretation process is the lack of consistency in marker results for any neoplasm.2-4 A certain marker may be positive (or negative) for a certain neoplasm in most of the cases. However, exceptions often are seen. For this reason, an absolute diagnostic
Am J Clin Pathol 2000;113:95106
95

American Society of Clinical Pathologists

Nguyen et al / DATABASE FOR IMMUNOPHENOTYPING BY FLOW CYTOMETRY

pattern usually is not available for any neoplasm. Instead, the diagnostic approach is to seek a neoplasm that closely matches the given marker results.2 Since the immunophenotyping pattern of hematopoietic neoplasms can be described easily in terms of the presence or absence of markers contained in separate fields, a database is a logical approach to facilitate the interpretation of marker results. In addition, a decision-support system that readily displays the attributes associated with a specific neoplasm would be convenient and useful to pathologists who encounter the more uncommon diseases from time to time. Several computer programs have been developed for the analysis of immunophenotyping results in malignant lymphoma and leukemia. 6-10 In the present study, we designed and evaluated a relational database for interpretation of immunophenotyping by flow cytometry. The database is designed to help the user follow closely how the suggested diagnoses are determined. The database also permits the user to modify the contents of the database to suit individual preferences.

Materials and Methods

The first phase of our study involved the development of the database, CD Marker. In the second phase, the database was validated by using previously interpreted cases at our affiliated hospitals. Relational Database A database is a computer program that stores and retrieves information based on a defined relationship.11 By using a database, one can organize data according to subjects so that the data will be easy to track and verify. Data in a database are stored in tables. A table is a collection of data on a particular subject. Data in a table are presented in columns (called fields) and rows (called records). Typical examples of fields and records in a database for hematopoietic neoplasms are disease attribute (such as immunologic marker result) and type of neoplasm, respectively. In a relational database, one also can store information on how different subjects are related to facilitate compiling the associated data. This information is essentially an expression that links records in one table with records in another. Analysis of data in a database can be performed conveniently with query.11 A query is a description of sets of records that meet certain criteria defined by the user. In a database for hematopoietic neoplasms, for example, a query may be used to look for neoplasms with marker results that match those of the case under consideration. Relational databases with comprehensive content and efficient query mechanisms can perform effectively as deci96

sion-support systems to help users solve complicated problems in laboratory diagnosis.12-14 Such systems have 2 major elements: the development environment and the consultation environment Figure 1. The development environment is used by the database builder to construct the components and to enter information into the knowledge base. The consultation environment is accessed by the user to obtain recommendations. The following components are seen in our database: 1. Knowledge base. This component contains the knowledge required for formulating and solving problems. It contains facts in the domain area and rules that direct the use of facts to diagnose specific disorders. The diagnostic criteria are the facts, also known as attributes, that are necessary to confirm a certain disorder. Potential sources of knowledge include human experts and literature. 2. Workplace. This component is an area of the computers working memory for the description of a current problem, as specified by the input data. The workplace also records intermediate conclusions. 3. Inference engine. The brain of the database, this component provides the method for query by using information in the knowledge base and in the workplace to formulate conclusions. 4. User interface. This component allows communication between the user and the database. The user uses this interface to input data (positive and negative attributes found in a patient), to obtain the results, and also to revise contents of the database as needed. This communication interface usually is in a graphics format for ease of use (graphic user interface). Database CD Marker CD Marker, a relational database for interpretation of immunophenotyping by flow cytometry, was implemented in Microsoft Access 97 (Microsoft, Redmond, WA) and was run on an IBM-compatible computer under Microsoft Windows 95. The inference engine of CD Marker was built with Microsoft Visual Basic for Application language. This inference engine uses a backward-chaining search strategy to draw conclusions. 15,16 A total of 33 hematopoietic neoplasms were included in the database Table 1. The diagnostic criteria for different neoplasms were based on the pattern of immunologic marker results.1-4 The marker result was designated as positive (or negative) for a neoplasm if more than 80% of the cases were found to be positive (or negative) for that marker. 4 A total of 42 immunologic markers were included in CD Marker Table 2. The current marker panel includes only the markers deemed to be used most commonly. As other markers become more extensively used with their added value in diagnosis, they will be incorporated into the database. A complete listing of immunophenotype for each neoplasm is
American Society of Clinical Pathologists

Am J Clin Pathol 2000;113:95106

Hematopathology / ORIGINAL ARTICLE

Table 1 Hematopoietic Neoplasms in CD Marker Database

Consultation Development Follicular small cleaved cell lymphoma Mantle cell lymphoma Large B-cell lymphoma Mediastinal B-cell lymphoma B-cell lymphoma, unclassifiable, mixed small and large cells Lymphoplasmacytoid lymphoma Marginal cell lymphoma Burkitt lymphoma/ALL, L3 Splenic lymphoma with villous lymphocytes Peripheral T-cell lymphoma Lymphoblastic lymphoma (T cell) Thymoma Chronic lymphocytic leukemia (B cell)/small lymphocytic lymphoma Chronic lymphocytic leukemia (T cell) Prolymphocytic leukemia (B cell) Hairy cell leukemia Szary syndrome/mycosis fungoides Adult T-cell leukemia/lymphoma ALL (T-cell precursor) Acute myeloblastic leukemia without maturation, M1 Acute myeloblastic leukemia with maturation, M2 Acute promyelocytic leukemia, M3 Acute myelomonocytic leukemia, M4 Acute monoblastic leukemia, M5 Acute erythroleukemia, M6 Acute megakaryoblastic leukemia, M7 Biphenotypic acute leukemia, AML and T-cell ALL Biphenotypic acute leukemia, AML and B-cell precursor ALL Large granular lymphocyte leukemia, NK cell Large granular lymphoproliferative disorder, T cell ALL (early-B precursor) ALL (CALLA) ALL (Pre-B)
ALL, acute lymphoblastic leukemia; AML, acute myeloblastic leukemia; NK, natural killer; CALLA, common acute lymphoblastic leukemia antigen.

User

Knowledge base: facts, rules

User interface

Workplace

Inference engine

Figure 1 Structure of a relational database used as a decision-support system.

included in Table 3. Figure 2 shows the main menu of the database with 3 modules: 1. Consultation for differential diagnosis 2. Look up or revise a disorder in the database 3. Look up or revise a marker in the database The graphic user interface for the differential diagnosis module is shown in Figure 3. All the data that are available on marker results should be entered for the case under consideration. Lack of information in certain data fields does not prevent CD Marker from processing the data. However, the accuracy of the suggested diagnosis would be compromised if results of important markers were left out. A list of differential diagnoses is provided by CD Marker with each set of input data. The differential diagnoses have an assigned value of matching factor (MF). The MF value for a neoplasm reflects how well its immunophenotyping pattern matches the marker data for a given case. This factor is defined as (Equation 1): MF = M/(M + N) where MF is the matching factor for a particular neoplasm (0 MF 1); M, the number of attributes of a neoplasm that match the input data; and N, the number of attributes of a neoplasm that do not match the input data. Note that the value of MF, as defined by Equation 1, reflects only the similarity between the attributes of a neoplasm and the available data. A high value of MF for a neoplasm, such as 1, does not exclude the possibility that more data input with increased value of N actually may decrease its MF value. To demonstrate how the knowledge base of CD Marker is implemented, the diagnostic criteria for chronic lymphocytic leukemia/small lymphocytic lymphoma (CLL/SLL)2-4 are given as an example:
American Society of Clinical Pathologists

Table 2 Immunologic Markers in CD Marker Database

CD1 CD2 CD3 CD4 CD5 CD7 CD8 CD10 CD11b CD11c CD13 CD14 CD15 CD16 CD19 CD20 CD21 CD22 CD23 CD24 CD25 CD33 CD38 CD41 CD42 CD43 CD45 CD56 CD57 CD61 CD71 CD77 CD79a CD103 HLA-DR sIg cIg PC-1 TdT FMC-7 Glycophorin A Cytokeratin

cIg, cytoplasmic immunoglobulin; sIg, surface immunoglobulin; TdT, terminal deoxynucleotidyl transferase.

Am J Clin Pathol 2000;113:95106

97

Nguyen et al / DATABASE FOR IMMUNOPHENOTYPING BY FLOW CYTOMETRY

Table 3 Immunophenotype for Each Disorder1-4

Disorder Follicular small cleaved cell lymphoma Mantle cell lymphoma Large B-cell lymphoma Mediastinal B-cell lymphoma B-cell lymphoma, unclassifiable, mixed small and large cells Lymphoplasmacytoid lymphoma Marginal cell lymphoma Burkitt lymphoma/ALL, L3 Splenic lymphoma with villous lymphocytes Peripheral T-cell lymphoma Lymphoblastic lymphoma (T cell) Thymoma Chronic lymphocytic leukemia (B cell)/small lymphocytic lymphoma Chronic lymphocytic leukemia (T cell) Prolymphocytic leukemia (B cell) Hairy cell leukemia Szary syndrome/mycosis fungoides Adult T-cell leukemia/lymphoma ALL (T-cell precursor) Acute myeloblastic leukemia without maturation, M1 Acute myeloblastic leukemia with maturation, M2 Acute promyelocytic leukemia, M3 Acute myelomonocytic leukemia, M4 Acute monoblastic leukemia, M5 Acute erythroleukemia, M6 Acute megakaryoblastic leukemia, M7 Biphenotypic acute leukemia, AML and T-cell ALL Biphenotypic acute leukemia, AML and B-cell precursor ALL Large granular lymphocyte leukemia, NK cell Large granular lymphoproliferative disorder, T cell ALL (early-B precursor) ALL (CALLA) ALL (pre-B) Positive CD10, CD19, CD20, CD21, CD22, CD24, HLA-DR, sIg CD5, CD19, CD20, CD22, CD24, CD43, HLA-DR, sIg CD19, CD20, CD22, CD79a, HLA-DR CD19, CD20, CD22, CD79a, PC-1 CD19, CD20, CD22, CD79a, HLA-DR, sIg CD19, CD20, CD22, CD79a, sIg, cIg CD19, CD20, CD22, CD79a, sIg CD10, CD19, CD20, CD22, CD77 , CD79a, HLA-DR, sIg CD19, CD20, CD22, CD79a, HLA-DR, sIg CD2, CD3, CD5, CD7 CD2, CD3, CD5, CD7 CD38, CD71, TdT , TdT, Cytokeratin CD5, CD19, CD20, CD21, CD23, CD24, CD43, CD79a, HLA-DR, sIg CD2, CD3, CD5, CD7 CD19, CD20, CD22, HLA-DR, sIg, FMC-7 CD11c, CD19, CD20, CD22, CD25, CD79a, CD103, sIg, FMC-7 CD2, CD3, CD4, CD5 CD2, CD3, CD4, CD5, CD25, CD38, CD71, HLA-DR CD3, CD7 TdT , CD13, CD33, HLA-DR CD13, CD15, CD33, HLA-DR CD13, CD15, CD33 CD11b, CD13, CD14, CD15, CD33, HLA-DR CD11b, CD11c, CD13, CD14, CD15, CD33, HLA-DR CD71, glycophorin A CD33, CD41, CD42, CD61 CD3, CD7 CD13, CD33, TdT , CD13, CD19, CD22, CD33, CD79a, HLA-DR, TdT CD2, CD5, CD16, CD56, HLA-DR CD2, CD3, CD16, CD56, CD57 CD19, CD22, CD79a, HLA-DR, TdT CD10, CD19, CD22, CD79a, HLA-DR, TdT CD10, CD19, CD22, CD79a, HLA-DR, cIg, TdT Negative CD2, CD3, CD4, CD5, CD7 CD8, , CD11c, CD23, CD25, CD43, cIg CD11c, CD23, cIg, CD5/CD19 or CD5/CD20 CD2, CD3, CD5, CD7 CD2, CD3, CD5, CD7 CD10, CD15, , CD21, sIg, TdT CD2, CD3, CD5, CD7 CD5, CD10, CD23 CD5, CD10, CD23 CD5, CD23, TdT CD5, CD10, CD23, CD25 CD19, CD20, CD22, cytokeratin CD15, CD20, cytokeratin CD4, CD8, CD45 CD2, CD3, CD4, CD7 CD8, CD10, , CD25, FMC-7 CD5/CD19 or , CD5/CD20 CD16, CD25, CD56, CD57 TdT , CD2, CD3, CD4, CD7 CD8, CD10, , CD11c, CD25, TdT CD2, CD3, CD4, CD5, CD7 CD8, , CD10, CD23 CD1, CD7 CD8, CD10, CD11c, CD16, , CD19, CD20, CD22, CD25, CD56, CD57 sIg, TdT, cytokeratin , CD1, CD7 CD8, CD10, CD11c, CD16, , CD19, CD20, CD22, CD56, CD57 sIg, , TdT, cytokeratin CD10, CD19, CD20, CD22, HLA-DR, cytokeratin CD2, CD3, CD5, CD7 CD11b, CD14, , CD15, CD41, CD42, CD61, CD71, sIg, glycophorin A CD2, CD3, CD5, CD7 CD11b, CD14, , CD41, CD42, CD61, CD71, sIg, glycophorin A CD2, CD3, CD5, CD7 CD11b, CD14, , CD41, CD42, CD61, CD71, HLA-DR, sIg, glycophorin A CD2, CD3, CD5, CD7 CD41, CD42, , CD61, CD71, sIg, glycophorin A CD2, CD3, CD5, CD7 CD41, CD42, , CD61, CD71, sIg, glycophorin A CD2, CD3, CD5, CD7 CD11b, CD14, , CD41, CD42, CD61, sIg CD2, CD3, CD5, CD7 CD11b, CD13, , CD14, CD15, CD71, sIg, glycophorin A CD10, CD19, CD20, CD22, cytokeratin CD3, CD7 sIg , CD3, CD4 CD25, TdT CD3, CD4, CD5, CD7 CD8, CD10, sIg, , cIg CD3, CD4, CD5, CD7 CD8, sIg, cIg , CD3, CD4, CD5, CD7 CD8, sIg ,

ALL, acute lymphoblastic leukemia; AML, acute myeloblastic leukemia; CALLA, common acute lymphoblastic leukemia antigen; cIg, cytoplasmic immunoglobulin; NK, natural killer; sIg, surface immunoglobulin; TdT, terminal deoxynucleotidyl transferase.

98

Am J Clin Pathol 2000;113:95106

American Society of Clinical Pathologists

Hematopathology / ORIGINAL ARTICLE

Figure 2 The main menu of CD Marker.

Figure 3 Graphic user interface to enter data. cIg, cytoplasmic immunoglobulin; sIg, surface immunoglobulin; TdT, terminal deoxynucleotidyl transferase.

1. The malignant cells are positive for the following markers: CD5, CD19, CD20, CD21, CD23, CD24, CD43, CD79a, HLA-DR, surface immunoglobulin, and coexpression of CD5 and CD19 or CD20. 2. The malignant cells are negative for the following markers: CD2, CD3, CD4, CD7, CD8, CD10, CD25, and FMC7. A summary of diagnostic attributes for CLL/SLL is shown in Figure 4. Note that some marker results are left blank in this set of attributes. Only the markers considered to have an important role in the differential diagnosis are included. This design helps to minimize the computing time in the inference process. When marker results of a given case are entered, they are processed by the database inference engine, and a list of differential diagnoses will be displayed. These diagnoses are listed with their associated MF value. A demonstration of a case with CLL/SLL illustrates of how CD Marker can be used for interpretation of immunophenotyping results and how its search mechanism works. Figure 5 shows the marker data available for the patient sample. CD Marker attempts to match this set of data with the diagnostic attributes of 33 hematopoietic neoplasms in the knowledge base. The available data matched the following attributes of CLL/SLL: 1. Positive for CD5, CD19, CD20, CD23, HLA-DR, surface immunoglobulin, and coexpression of CD5 and CD19. 2. Negative for: CD3, CD4, CD7, CD8, CD10, and CD25. The total number of attributes of CLL/SLL that matched the input data was 13 (7 positive results and 6 negative results). This number was represented by the variable M in Equation 1 (M = 13). None of the attributes of CLL/SLL was in conflict with the input data (N = 0). Note that the following input data did not have a corresponding attribute
American Society of Clinical Pathologists

Figure 4 Immunophenotyping for chronic lymphocytic leukemia (B cell)/small lymphocytic lymphoma. cIg, cytoplasmic immunoglobulin; Neg, negative; sIg, surface immunoglobulin; TdT, terminal deoxynucleotidyl transferase; TRAP tartrate-resistant acid phosphatase. ,

for CLL/SLL in the knowledge base: CD11c, CD14, CD16, CD22, CD45, glycophorin A, and cytokeratin. These input data had no effect on the ranking of CLL/SLL since they were not included as part of the calculation for its MF. Similarly, the following attributes in the knowledge base without corresponding input data had no effect on the MF value for CLL/SLL: CD2, CD21, CD24, CD43, CD79a, and FMC7. The intentional exclusion of input data without corresponding attributes in the database (or attributes without corresponding input data) in calculating MF serves the important purpose of maintaining a flexible design for the knowledge base and for the data input panel. Since different
Am J Clin Pathol 2000;113:95106
99

Nguyen et al / DATABASE FOR IMMUNOPHENOTYPING BY FLOW CYTOMETRY

Figure 5 Input data for a case of chronic lymphocytic leukemia (CLL). cIg, cytoplasmic immunoglobulin; sIg, surface immunoglobulin; TdT, terminal deoxynucleotidyl transferase.

Figure 6 Output for the demonstration case of chronic lymphocytic leukemia (CLL). ALL, acute lymphoblastic leukemia; AML, acute myeloblastic leukemia; NK, natural killer.

flow cytometry laboratories may use different markers for immunophenotyping and various studies on marker pattern of neoplasms have used different marker panels, an absolute requirement of certain markers in the interpretation process would be too stringent to yield reasonable matches.5 The MF value for CLL/SLL at this point was: MF = 13 / (13 + 0) = 1 After CD Marker calculated the MF value for all the remaining 32 hematopoietic neoplasms in the knowledge base and ranked them accordingly, it listed the following leading diagnoses: 1. CLL/SLL; MF = 1 2. Prolymphocytic leukemia (B-cell); MF = 1 3. Mantle cell lymphoma; MF = 0.89 4. Mixed cell lymphoma; MF = 0.88 5. Large B-cell lymphoma; MF = 0.86 CLL/SLL and prolymphocytic leukemia had the same MF value (MF = 1) with the given input data. A second criterion, the difference between the matched attributes and the unmatched attributes for a neoplasm (M N), is used to refine the ranking process for neoplasms with the same MF value. With this second criterion, CLL/SLL was ranked as the leading diagnosis (MF = 1, M N = 13), followed by prolymphocytic leukemia (MF = 1, M N = 12). Figure 6 shows the list of differential diagnoses with all the calculation results. The mechanism of searching from neoplasms in the database to the input data for the best matches represents a strategy known as backward-chaining search.11,15,16 This demonstration shows the open-ended format of the data input. The data panel consists of many immunologic markers, some of which may not be part of routine testing in a particular laboratory. Consequently, the actual data input for a case are unlikely to account for all the
100

markers in the data panel. However, the availability of essential data would influence the accuracy of ranking by CD Marker. In the demonstration case of CLL/SLL, the ranking results would be different if the result for CD5 and CD23 were not available. In this scenario Figure 7, prolymphocytic leukemia would be the leading diagnosis (MF = 1, M N = 12), followed by CLL/SLL (MF = 1, M N = 11). For comparison, the diagnostic criteria for prolymphocytic leukemia2-4 are shown in Figure 8. The critical role of the interpreting pathologist cannot be overemphasized. CD Marker is useful only for suggesting a list of differential diagnoses. The pathologist must establish the final diagnosis by correlating the histologic findings of the case with the immunophenotyping results. The immunologic marker pattern of neoplasms in the list of differential diagnoses can be reviewed during the interpretation process by using the Look up a disorder in database feature of CD Marker (Figures 4 and 8). As shown in the preceding demonstration, CD Marker was designed with a user-friendly interface. This graphic user interface was arranged such that the sequence of data entry, display of results, and review of marker pattern should be intuitive to the user. Besides the essential features shown in the demonstration, CD Marker also has been designed to include the following features: 1. What If reruns can be performed whenever the user edits one or more of the input data. This editing may be in the form of changing the value of the data (positive to negative or vice versa), deleting the input data, or adding additional data. The option for reruns enhances the flexibility of CD Marker for evaluating data that may not be clear-cut, that is, when the marker results are borderline. Without requiring different sets of data entry, the What If
American Society of Clinical Pathologists

Am J Clin Pathol 2000;113:95106

Hematopathology / ORIGINAL ARTICLE

reruns offer the user a convenient way to consider all the potential diagnoses based on laboratory data that may be subject to errors for various reasons. 2. The user has the option to revise the database contents to fit individual needs. In other words, the immunologic marker pattern of neoplasms in CD Marker can be modified to the users preference. CD Marker has default tables that together form the knowledge base of the database. This knowledge base has been validated in the present study (see later sections on methods of validation and validation results). When the user edits the database tables, the tables are updated automatically to reflect the revisions made by the user. The revised tables will be used by CD Marker in subsequent runs. Since the revised tables contain information that has not been validated in our study, the user must be aware of the need to conduct individual validation studies to ensure that CD Marker performs adequately after each revision. CD Marker offers the user the option of selecting the default tables or the revised tables before each run. CD Marker makes this option possible by saving the default tables and the revised tables before any revision by the user. 3. The contents on the computer screen at any time during a session with CD Marker can be printed for hard copy report. 4. Online HELP instructions are available for using CD Marker. Method for Validating Database CD Marker In the second phase of our project, we used 92 cases involving various hematopoietic neoplasms to assess the performance of CD Marker. These were patient cases at Lyndon B. Johnson Hospital and Ben Taub Hospital (Harris County Hospital District, Houston, TX). Flow cytometry studies for these cases were performed between January 1995 and December 1996. Data for these 92 cases were retrieved retrospectively, and immunophenotyping data were tested on CD Marker. The number of cases for each disorder is shown in Table 4. The final diagnosis for each case was established previously by histologic findings and correlation with flow cytometry results. The final diagnosis was documented in surgical pathology reports, including bone marrow reports. Data entry for each case for CD Marker included only marker results that were available in the flow cytometry laboratory at Ben Taub Hospital at the time of initial presentation. Only definitive marker results (positive or negative) in each case were used in the validation process. Equivocal results were not used owing to their lack of contribution to the validation results. To avoid potential bias in the design of CD Marker, its knowledge base was developed by one of us (A.N.D.N.) who had not interpreted the cases previously. Immunophenotyping by flow cytometry was performed on a FACScan (Becton Dickinson, Mountain View, CA) as
American Society of Clinical Pathologists

Figure 7 Output for the demonstration case without data input for CD5 and CD23.

Figure 8 Immunophenotyping for prolymphocytic leukemia. cIg, cytoplasmic immunoglobulin; Glyco, glycophorin; Pos, positive; sIg, surface immunoglobulin; TdT, terminal deoxynucleotidyl transferase; TRAP tartrate, resistant acid phosphatase.

described previously.1 The specimens in our cases included bone marrow, lymph node, spleen, body fluid, and extranodal hematopoietic tumors. For a CD Marker session to be considered successful, the final diagnosis for a given case must be included in the list of 5 differential diagnoses suggested by CD Marker.

Results
Table 5 shows the preliminary results for all the 92 cases used in this validation process with the accompanying information: ranking of the final diagnosis by CD Marker, the number of cases in each ranking category and the
Am J Clin Pathol 2000;113:95106
101

Nguyen et al / DATABASE FOR IMMUNOPHENOTYPING BY FLOW CYTOMETRY

Table 4 Number of Cases Used in Validation (N = 92)

Disorder Follicular small cleaved cell lymphoma Mantle cell lymphoma Large B-cell lymphoma Mediastinal B-cell lymphoma B-cell lymphoma, unclassifiable, mixed small and large cells Lymphoplasmacytoid lymphoma Marginal cell lymphoma Burkitt lymphoma/ALL, L3 Splenic lymphoma with villous lymphocytes Peripheral T-cell lymphoma Lymphoblastic lymphoma (T cell) Thymoma Chronic lymphocytic leukemia (B cell)/small lymphocytic lymphoma Chronic lymphocytic leukemia (T cell) Prolymphocytic leukemia (B cell) Hairy cell leukemia Szary syndrome/mycosis fungoides Adult T-cell leukemia/lymphoma ALL (T-cell precursor) Acute myeloblastic leukemia without maturation, M1 Acute myeloblastic leukemia with maturation, M2 Acute promyelocytic leukemia, M3 Acute myelomonocytic leukemia, M4 Acute monoblastic leukemia, M5 Acute erythroleukemia, M6 Acute megakaryoblastic leukemia, M7 Biphenotypic acute leukemia, AML and T-cell ALL Biphenotypic acute leukemia, AML and B-cell precursor ALL Large granular lymphocyte leukemia, NK cell Large granular lymphoproliferative disorder, T cell ALL (early-B precursor) ALL (CALLA) ALL (pre-B)
ALL, acute lymphoblastic leukemia; AML, acute myeloblastic leukemia; CALLA, common acute lymphoblastic leukemia antigen.

No. of Cases 5 2 5 2 2 2 3 3 4 5 2 2 7 2 2 4 2 2 2 3 4 2 2 2 2 2 2 2 2 2 2 3 4

Table 5 Summary of Validation Results

Preliminary Accumulated Percentage Final Accumulated Percentage

Ranking by CD Marker Differential diagnosis First Second Third Fourth Fifth Lower ranking Total

No. (%) of Cases

No. (%) of Cases

36 (39) 23 (25) 11 (12) 10 (11) 2 (2) 10 (11) 92 (100)

39 64 76 87 89

39 (42) 23 (25) 12 (13) 10 (11) 2 (2) 6 (6) 92 (99)*

42 67 80 91 93

*Total does not equal 100 because of rounding.

percentage and accumulated percentage of cases in each ranking category. The validation results showed a success rate of 89%. This success rate means that in 89% of the cases, the final diagnosis was included in the list of 5 differential diagnoses generated by CD Marker. In 11% of the cases, the final diagnosis was not included in this list for the following reasons: 1. Inadequate data input: Small lymphocytic lymphoma
102

(SLL) in case 25 was not ranked by CD Marker owing to lack of a light chain restriction. Review of flow cytometric data revealed a B-cell subpopulation with a definite lambda light chain restriction. SLL subsequently was ranked first for this case by CD Marker. 2. Incorrect final diagnosis: Cases 1 and 11 were diagnosed as SLL, which was not ranked by CD Marker. Review of microscopic slides from these cases by a pathologist in
American Society of Clinical Pathologists

Am J Clin Pathol 2000;113:95106

Hematopathology / ORIGINAL ARTICLE

our group (E.I.B.) indicated that the diagnosis of these cases should be revised as mantle cell lymphoma (case 1) and large cell lymphoma (case 11). In fact, mantle cell lymphoma was ranked first and large cell lymphoma was ranked third by CD Marker for cases 1 and 11, respectively. 3. Limitations of CD Marker knowledge base: Case 9 was diagnosed as acute lymphoblastic leukemia (ALL), early preB cell type. This diagnosis was not ranked by CD Marker owing to a lack of B-cell ALL subtypes in the database. The general classification of B-cell precursor ALL included insufficient diagnostic criteria for early pre-B cell ALL. A revision was made for CD Marker to include all B-cell ALL subtypes. The correct diagnosis of early pre-B cell ALL subsequently was ranked first after this revision. Three cases of T-cell lymphoma (cases 37, 67, and 68) were not ranked by CD Marker. This deficiency was due to an intrinsic limitation of CD Marker in handling certain cases of T-cell malignant neoplasms. Aberrant loss of T-cell antigens is a characteristic finding in T-cell malignant neoplasms.2-4 However, a suitable inference mechanism has not been developed to detect such a manifestation. The difficulty in designing an algorithm for such detection lies in the random distribution of T-cell markers, which makes it impossible to program the values of diagnostic attributes into the database. Despite this shortcoming, a considerable number of T-cell cases were ranked successfully by CD Marker (11 [78%] of 14 T-cell cases). 4. Unusual marker results: One was a case of CD5+ large cell lymphoma (case 38). Review of microscopic slides from this case indicated that the histologic and immunophenotyping results were suggestive of Richter syndrome. However, the available clinical information was insufficient to confirm this diagnosis. Another was a case of HLADRnegative large cell lymphoma (case 14). The reason for negative HLA-DR result was unknown but most likely was technical error. Last was a case of CD5 SLL (case 31). Review of microscopic slides from this case confirmed the final diagnosis of SLL. In the light of the new ranking by CD Marker for the cases under consideration, our validation results were revised to reflect the successful ranking for 4 additional cases (1, 9, 11, and 25). The revised results, summarized in Table 5, showed successful ranking in 93% of the cases (the final diagnosis included in the list of top 5 differential diagnoses). Furthermore, the final diagnosis was ranked first in 39 (42%) of 92 cases.

Discussion
A number of computer programs have been developed to facilitate interpretation of immunophenotyping for hematopoietic neoplasms by flow cytometry.6-10 Different
American Society of Clinical Pathologists

approaches have been attempted, including rule-based systems, cluster analysis, semantic networks, and various mathematical algorithms. Alvey et al6 designed a rule-based production system for the diagnosis of acute and chronic leukemia. Suggested diagnoses by the program are qualified as definite, probable, compatible, or possible. This program gave correct conclusions for 400 cases after 4 iterations in a retrospective study, along with a summary of its reasoning process and suggestions for further testing if necessary. This program was developed specifically for leukemia and, therefore, is limited to only a subset of hematopoietic neoplasms. Petrovecki et al7 developed a mathematical algorithm based on matrix algebra. Their program calculates compatibility scores to rank potential diagnoses and covers a wide spectrum of hematopoietic neoplasms. The program was tested with 58 cases in a retrospective study and suggested the correct diagnosis as 1 of the top 4 choices in 93% of the cases. However, the correct diagnosis was ranked first in only 29% of the cases. Verwer et al8 used cluster analysis to automatically assign cell lineage to acute leukemic cells during the analytic phase of flow cytometric study. In this method, multidimensional data from data files in list mode are divided into groups using the nearest-neighbor algorithm. The position of the cellular groups is analyzed by a criteria table. The program then classifies the groups into appropriate cell lineage. This technique of automated gating in the analytic phase of flow cytometric study is expected to reduce technical errors commonly seen with the conventional way of manual gating. This program, however, is limited to leukemia and was tested only on a small number of clinical cases. Diamond et al9 developed a relational database, known as Professor Fidelio, that matches immunophenotyping patterns of hematopoietic neoplasms with the input data. This database also covers other types of diagnostic data, including cellular morphology, cytochemistry, DNA content and proliferative activity. The program was validated successfully in 300 (80.0%) of 366 test cases. This is one of the programs that offer the most comprehensive option for data input. The only drawback is the limited number of diagnoses in the database, with only 12 general categories of hematopoietic neoplasms. Thews et al10 used semantic networks to model a hierarchy of hematopoietic cells and their occurrence in neoplasms. Their program is limited to blood and bone marrow samples of a few specific types of neoplasms: acute myeloblastic leukemia, B-cell ALL, and B-cell lymphoma. The validation study showed an impressive success rate of 97.3% of the cases (616/633). Although several programs have been introduced to facilitate computer-assisted interpretation of flow cytometric
Am J Clin Pathol 2000;113:95106
103

Nguyen et al / DATABASE FOR IMMUNOPHENOTYPING BY FLOW CYTOMETRY

data, we believe that CD Marker possesses several unique features not found in other similar systems. The diagnostic criteria and the associated inference process we describe demonstrate a clear and understandable way of knowledge representation for marker interpretation. The inference process is no longer a black box with all the complex algorithms incomprehensible to people outside the fields of mathematics and artificial intelligence. This advantage, coupled with the user-friendly interface in a Windows environment, potentially can increase the acceptance of computer-assisted interpretation by pathologists who analyze flow cytometric data. Confidence in using CD Marker also is enhanced by the option to modify its database contents to suit individual needs. The database is comprehensive and also can be revised easily to incorporate more disease entities as needed. Furthermore, our program can run on an average personal computer with a minimum requirement for hardware and software (486 microprocessor with 16 megabytes of RAM, Microsoft Windows 95). The run-time version of CD Marker does not require that Microsoft-Access database be installed in the users computer. Recently, we posted CD Marker on the World Wide Web to make it available to any user who needs access to this database. The URL for our Web page is as follows: http://dpalm.med.uth.tmc.edu/faculty/bios/nguyen/nguyen.html The only requirement to run this database directly on the Web is the availability of the Microsoft-Access 97 program in the users computer. Copyright is applicable to the contents of CD Marker to protect our ownership. Despite the usefulness of CD Marker, the following constraints are inherent in its use. 1. This database is designed for pathologists who are not subspecialized in hematopathology and for pathology residents-in-training. It simply provides the differential diagnosis for a given set of marker expressions without regard to morphologic features, clinical manifestations, or response to therapy. Therefore, the differential diagnosis is designed to be broad. Just like trained hematopathologists, nonhematopathologists must correlate the immunophenotypic findings with the morphologic characteristics of the neoplasm, the most important basis of diagnosis and classification. 2. The user must have a functional knowledge of hematopoietic disorders to be able to use CD Marker effectively because this database serves only as a search tool to aid the user in making a diagnosis. The technical skills to perform the laboratory procedures and the experience needed to accurately gate the cellular populations are critical in the diagnostic process. CD Marker can generate a list of differential diagnoses in most cases if adequate data are input. The interpreting pathologist then can quickly compare the patients laboratory data with the marker patterns available from the CD Marker display module and make the appro104

priate diagnosis. It cannot be overemphasized that human judgment is the most important element in finalizing the diagnosis. The number and complexity of hematopoietic neoplasms require that the differential diagnoses suggested by CD Marker be reviewed before making a final diagnosis. We believe that the pathologists clinical judgment can be enhanced with the information from CD Marker to yield an accurate diagnosis. 3. The current version of CD Marker is deficient in handling some cases of T-cell malignant neoplasm owing to the difficulty designing an algorithm for detection of the random loss of T-cell antigens as discussed earlier. We are working on several approaches to alleviate this shortcoming and expect to implement a new technique to handle T-cell malignant neoplasms more effectively in future versions of CD Marker. 4. Not all the commercially available markers were used in our laboratory. Subsequently, our validation results do not represent a maximal accuracy level that would have been achieved if all the available markers had been used. Further research is needed to analyze results obtained from different laboratories using markers not available in our laboratory. 5. CD Marker would not be useful in the diagnosis of neoplasms that traditionally have not been shown to benefit from flow cytometric immunophenotyping.1-4 These include the following: (1) Hodgkin disease; (2) multiple myeloma; (3) neoplasms with common immunophenotypes: mucosaassociated lymphoid tissue lymphoma, marginal cell lymphoma, monocytoid B-cell lymphoma, lymphoplasmacytoid lymphoma, and many cases of large B-cell lymphoma. These neoplasms have the same immunophenotype CD19+/CD20+, CD5, CD10, and CD23. Also, follicular center cell lymphoma and Burkitt lymphoma have the same immunophenotype CD19+/CD20+, CD5, CD10+, and CD23. (4) Atypical immunophenotypes: CD5 CLL/SLL, CD23 CLL/SLL, CD23+ mantle cell lymphoma, CD10 follicular center cell lymphoma, CD5+ mucosa-associated lymphoid tissue lymphoma, and CD5+ marginal cell lymphoma; (5) T-cell lymphomas without loss of T-cellassociated antigens, which may be misdiagnosed as benign T-cell hyperplasia; (6) aberrant expression of Tcell antigens on B-cell lymphoma, such as CD2+ CLL/SLL, CD8+ CLL/SLL, and CD8+ mantle cell lymphoma; (7) polyclonal posttransplant B-cell lymphoproliferative disorders; (8) CD34 acute leukemia; (9) terminal deoxynucleotidyl transferasenegative acute leukemia; (10) acute promyelocytic leukemia without a diagnostic immunophenotypic profile, such as CD34, HLA-DRnegative, CD13+, CD33+, CD41a, and glycophorin Anegative. The immunophenotype merely reflects the presence of a myeloid cell population, and the same immunophenotype may be obtained in cases of chronic myelocytic leukemia
American Society of Clinical Pathologists

Am J Clin Pathol 2000;113:95106

Hematopathology / ORIGINAL ARTICLE

and secondary leukocytosis. (11) T-cellrich B-cell lymphoma with a predominant population of reactive T lymphocytes, mostly CD4+ and in almost all instances do not show clonality by immunoglobulin light chain ratio because of a very small population of malignant large B lymphocytes. Such cases can be misdiagnosed easily as a reactive lymph node on flow cytometry data alone. The development of user-friendly software and the rapidly increasing number of physicians with skills and interests in computer-based applications create an ideal climate for the use of computers for diagnostic purposes.17-34 The number of database applications in laboratory medicine is expected to increase. Laboratory clinicians are in an ideal position to exploit database applications since they tend to be computer literate and have access to a large amount of patient data on laboratory computers.32 We found that CD Marker provides a convenient interactive tool to assist clinical personnel in diagnosing hematopoietic neoplasms by using flow cytometric data. This decision-support tool is available to a large number of users via the World Wide Web and also can be useful for independent study on immunophenotyping patterns of hematopoietic neoplasms. A logical extension of CD Marker would be to integrate the knowledge base into an existing laboratory flow cytometer. Data could then be retrieved directly from the instrument for use by the database, thus simplifying user interaction.
From the 1Departments of Pathology and Laboratory Medicine and 2Health Informatics, University of Texas Health Science Center at Houston, and the 3Department of Pathology, Baylor College of Medicine, Houston, TX. Address reprint requests to Dr Nguyen: Dept of Pathology and Laboratory Medicine, University of Texas Health Science Center at Houston, 6431 Fannin MSB 2.292, Houston, TX 77030. Acknowledgments: We thank Alex Buraruk, MT(ASCP), for consulting with us on technical matters during the course of this project; Donna Obermeier, for many helpful suggestions; and Robert Hunter, MD, PhD, for reviewing the manuscript with many insightful comments.

References
1. Keren DF. Flow Cytometry in Clinical Diagnosis. Chicago, IL: ASCP Press; 1989. 2. Sun T. Color Atlas and Text of Flow Cytometric Analysis of Hematologic Neoplasms. New York, NY: Igaku-Shoin; 1993. 3. Kjeldsberg C, ed. Practical Diagnosis of Hematologic Disorders. Chicago, IL: ASCP Press; 1995. 4. Harris NL, Jaffe ES, Stein H, et al. A revised EuropeanAmerican classification of lymphoid neoplasms: a proposal from the International Lymphoma Study Group. Blood. 1994;84:1361-1392.
American Society of Clinical Pathologists

5. Check W. Diagnosing leukemia, lymphoma: when laboratories go with flow analysis. CAP Today. November 1998:1-37. 6. Alvey PL, Preston NJ, Greaves MF. High performance for expert systems: a system for leukemia diagnosis. Med Inf. 1987;12:97-114. 7. Petrovecki M, Marusic M, Dezelic G. An algorithm for leukaemia immunophenotype pattern recognition. Med Inf. 1993;18:11-21. 8. Verwer B, Terstappen L. Automatic lineage assignment of acute leukemia by flow cytometry. Cytometry. 1993;14:862875. 9. Diamond LW, Nguyen DT, Andreeff M, et al. A knowledgebased system for the interpretation of flow cytometry data in leukemias and lymphomas. Cytometry. 1994;17:266-273. 10. Thews O, Thews A, Huber C, et al. Computer assisted interpretation of flow cytometry data in hematology. Cytometry. 1996;23:140-149. 11. Chignell M, Parsaye K. Expert Systems for Experts. New York, NY: John Wiley and Sons; 1988. 12. Ryan C. Cost reduction and QC software on a microbiology identification and susceptibility system [abstract]. Am Clin Lab. 1995:26. 13. Autio K, Kari A, Tikka H. Integration of knowledge-based system and database for identification of disturbances in fluid and electrolyte balance. Comput Methods Programs Biomed. 1991;34:201-209. 14. Connelly DP. Embedding expert systems in laboratory information systems. Am J Clin Pathol. 1990;94(suppl 1):7-14. 15. Buchanan BG, Shortliffe EH, eds. Rule-Based Expert Systems. Reading, MA: Addison-Wesley; 1984. 16. Turban E. Expert Systems and Applied Artificial Intelligence. New York, NY: Macmillan; 1992:73-144, 453-480. 17. Marquardt VC. Artificial intelligence and decision-support technology in the clinical laboratory. Lab Med. 1993;24: 777-782. 18. Barnett GO, Cimino JJ, Hupp JA. DXplain: an evolving diagnostic decision-support system. JAMA. 1987;258:67-74. 19. Forsstrom J, Nuutila P, Irjala K. Using the ID3 algorithm to find discrepant diagnoses from laboratory databases of thyroid patients. Med Decis Making. 1991;11:171-175. 20. Shortliffe EH. Computer programs to support clinical decision making. JAMA. 1987;258:61-66. 21. Groth T, Moden H. A knowledge-based system for real-time quality control and fault diagnosis of multitest analyzers. Comput Methods Programs Biomed. 1991;34:175-190. 22. Tischler AS, Martin MR. Generation of surgical pathology reports using a 5,000 word speech recognizer. Am J Clin Pathol. 1989;92(suppl 4, pt 1):S44-S47. 23. Tong DA. Weaning patients from mechanical ventilation: a knowledge-based system approach. Comput Methods Programs Biomed. 1991;35:267-278. 24. Weiss SM, Kulikowski CA, Galen RS. Representing experience in a computer program: the serum protein diagnostic program. J Clin Lab Automation. 1983;3:383-387. 25. Shifman MA. FABHELP: a rule-based consultation program for FAB classification of acute myeloid leukemia and myelodysplastic syndromes. Lab Med. 1991;22:639-643. 26. OConnor ML, McKinney T. The diagnosis of microcytic anemia by a rule-based expert system using VP-Expert. Arch Pathol Lab Med. 1989;113:985-988.
Am J Clin Pathol 2000;113:95106
105

Nguyen et al / DATABASE FOR IMMUNOPHENOTYPING BY FLOW CYTOMETRY

27. Spackman KA, Connelly DP. Knowledge-based systems in laboratory medicine and pathology. Arch Pathol Lab Med. 1987;111:116-119. 28. Bates JE, Bessman JD. Evaluation of BCDE, a microcomputer program to analyze automated blood counts and differentials. Am J Clin Pathol. 1987;88:314-323. 29. Blomberg DJ, Ladley JL, Fattu JM, et al. The use of an expert system in the clinical laboratory as an aid in the diagnosis of anemia. Am J Clin Pathol. 1987;87:608-613. 30. Nguyen DT, Diamond LW, Priolet G, et al. Expert system design in hematology diagnosis. Methods Inf Med. 1992; 31:82-89.

31. Nguyen AND, Hartwell EA, Milam JD. A rule-based expert system for laboratory diagnosis of hemoglobin disorders. Arch Pathol Lab Med. 1996;120:817-827. 32. Spackman KA, Connelly DP. Knowledge-based systems in laboratory medicine and pathology: a review and survey of the field. Arch Pathol Lab Med. 1987;111:116-119. 33. Siegel JD. Computerized diagnosis: implications for clinical education. Med Educ. 1988;22:47-54. 34. Nguyen AND, Uthman M. XPCOAG: a teaching software for laboratory diagnosis of coagulation disorders. Lab Med. 1994;25:399-401.

106

Am J Clin Pathol 2000;113:95106

American Society of Clinical Pathologists